Classification and Prediction Overview

Classification and Prediction Overview • • • • • • • • • • Data Mining Techniques: Classification and Prediction Mirek Riedewald Some slides based on presentations by Han/Kamber/Pei, Tan/Steinbach/Kumar, and Andrew Moore Introduction Decision Trees Statistical Decision Theory Nearest Neighbor Bayesian Classification Artificial Neural Networks Support Vector Machines (SVMs) Prediction Accuracy and Error Measures Ensemble Methods 2 Classification vs. Prediction Induction: Model Construction • Assumption: after data preparation, we have a data set where each record has attributes X1,…,Xn, and Y. • Goal: learn a function f:(X1,…,Xn)Y, then use this function to predict y for a given input record (x1,…,xn). Training Data Classification Algorithm – Classification: Y is a discrete attribute, called the class label • Usually a categorical attribute with small domain NAME M ike M ary B ill Jim D ave Anne – Prediction: Y is a continuous attribute • Called supervised learning, because true labels (Yvalues) are known for the initially provided data • Typical applications: credit approval, target marketing, medical diagnosis, fraud detection RANK YEARS TENURED A ssistan t P ro f 3 no A ssistan t P ro f 7 yes P ro fesso r 2 yes A sso ciate P ro f 7 yes A ssistan t P ro f 6 no A sso ciate P ro f 3 no Model (Function) IF rank = ‘professor’ OR years > 6 THEN tenured = ‘yes’ 3 Deduction: Using the Model Classification and Prediction Overview Model (Function) Test Data Unseen Data (Jeff, Professor, 4) NAME T om M erlisa G eorge Joseph RANK YEARS TENURED A ssistant P rof 2 no A ssociate P rof 7 no P rofessor 5 yes A ssistant P rof 7 yes 4 Tenured? 5 • • • • • • • • • • Introduction Decision Trees Statistical Decision Theory Bayesian Classification Artificial Neural Networks Support Vector Machines (SVMs) Nearest Neighbor Prediction Accuracy and Error Measures Ensemble Methods 6 1 Example of a Decision Tree Tid Refund Marital Status Taxable Income Cheat 1 Yes Single 125K No 2 No Married 100K No 3 No Single 70K No 4 Yes Married 120K No 5 No Divorced 95K Yes 6 No Married No 7 Yes Divorced 220K No 8 No Single 85K Yes 9 No Married 75K No 10 No Single 90K Yes 60K Another Example of Decision Tree Splitting Attributes Refund Yes No NO MarSt Married Single, Divorced TaxInc < 80K NO > 80K YES NO 10 Tid Refund Marital Status Taxable Income Cheat 1 Yes Single 125K No 2 No Married 100K No 3 No Single 70K No 4 Yes Married 120K No 5 No Divorced 95K Yes 6 No Married No 7 Yes Divorced 220K No 8 No Single 85K Yes 9 No Married 75K No 10 No Single 90K Yes 60K MarSt Married NO Single, Divorced Refund No Yes NO TaxInc < 80K > 80K YES NO There could be more than one tree that fits the same data! 10 Model: Decision Tree Training Data 7 8 Apply Model to Test Data Apply Model to Test Data Test Data Test Data Start from the root of tree. Refund Yes Refund No NO Yes MarSt Single, Divorced TaxInc < 80K MarSt Married Single, Divorced TaxInc NO > 80K < 80K YES NO No NO Married NO > 80K YES NO 9 10 Apply Model to Test Data Apply Model to Test Data Test Data Refund Yes Test Data Refund Marital Status Taxable Income Cheat No 80K Married ? Refund 10 No NO Yes MarSt Single, Divorced TaxInc < 80K NO TaxInc < 80K NO 11 80K Married ? MarSt NO > 80K No 10 Single, Divorced YES Taxable Income Cheat No NO Married Refund Marital Status Married NO > 80K YES 12 2 Apply Model to Test Data Apply Model to Test Data Test Data Refund Yes Test Data Refund Marital Status Taxable Income Cheat No 80K Married ? Refund 10 No NO Yes MarSt Single, Divorced TaxInc < 80K NO No 80K Married ? 10 MarSt TaxInc NO Assign Cheat to “No” Married Single, Divorced NO < 80K YES NO Taxable Income Cheat No Married > 80K Refund Marital Status > 80K YES NO 13 14 Decision Tree Induction Decision Boundary x2 • Basic greedy algorithm – – – – Top-down, recursive divide-and-conquer At start, all the training records are at the root Training records partitioned recursively based on split attributes Split attributes selected based on a heuristic or statistical measure (e.g., information gain) • Conditions for stopping partitioning – Pure node (all records belong to same class) – No remaining attributes for further partitioning 0.9 X1 < 0.43? 0.8 0.7 Yes NO MarSt 0.5 0.4 Yes :4 :0 0.1 Married Single, Divorced TaxInc < 80K NO > 80K No :0 :4 Yes No :0 :3 :4 :0 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 x1 1 Decision boundary = border between two neighboring regions of different classes. For trees that split on a single attribute at a time, the decision boundary is parallel to the axes. YES NO X2 < 0.33? X2 < 0.47? 0.2 No No 0.6 0.3 Refund Yes • Majority voting for classifying the leaf – No cases left 1 15 Oblique Decision Trees 16 How to Specify Split Condition? • Depends on attribute types – Nominal – Ordinal – Numeric (continuous) x+y<1 Class = + Class = • Depends on number of ways to split – 2-way split – Multi-way split • Test condition may involve multiple attributes • More expressive representation • Finding optimal test condition is computationally expensive 17 18 3 Splitting Nominal Attributes Splitting Ordinal Attributes • Multi-way split: use as many partitions as distinct values. • Multi-way split: Size Small Medium CarType Family Luxury • Binary split: Sports • Binary split: divides values into two subsets; need to find optimal partitioning. {Sports, Luxury} CarType {Family} OR Large {Family, Luxury} Size {Small, Medium} {Large} OR • What about this split? CarType {Sports} {Medium, Large} Size {Small} Size {Small, Large} {Medium} 19 Splitting Continuous Attributes 20 Splitting Continuous Attributes • Different options – Discretization to form an ordinal categorical attribute Taxable Income > 80K? • Static – discretize once at the beginning • Dynamic – ranges found by equal interval bucketing, equal frequency bucketing (percentiles), or clustering. Taxable Income? < 10K Yes > 80K No [10K,25K) – Binary Decision: (A < v) or (A  v) (i) Binary split [25K,50K) [50K,80K) (ii) Multi-way split • Consider all possible splits, choose best one 21 How to Determine Best Split How to Determine Best Split • Greedy approach: Before Splitting: 10 records of class 0, 10 records of class 1 Own Car? Yes Car Type? No Family C0: 6 C1: 4 C0: 4 C1: 6 C0: 1 C1: 3 Luxury C0: 8 C1: 0 – Nodes with homogeneous class distribution are preferred Student ID? c1 Sports C0: 1 C1: 7 C0: 1 C1: 0 ... 22 c10 C0: 1 C1: 0 c11 C0: 0 C1: 1 • Need a measure of node impurity: c20 ... C0: 0 C1: 1 Which test condition is the best? 23 C0: 5 C1: 5 C0: 9 C1: 1 Non-homogeneous, Homogeneous, High degree of impurity Low degree of impurity 24 4 Attribute Selection Measure: Information Gain Example • Select attribute with highest information gain • pi = probability that an arbitrary record in D belongs to class Ci, i= ,…, • Expected information (entropy) needed to classify a record m in D: Info(D)   pi log 2 ( pi ) • Predict if somebody will buy a computer Income Student Credit_rating Buys_computer • Given data set: Age High No Bad No 30 High  30 … High > 40 Medium > 40 Low Low > 40 31...40 Low  30 Medium  30 Low > 40 Medium  30 Medium 31...40 Medium 31...40 High > 40 Medium i 1 • Information needed after using attribute A to split D into v partitions D1,…, Dv: v |D | j Info A( D)   Info(D j ) j 1 | D | • Information gained by splitting on attribute A: Gain A (D)  Info(D)  Info A (D) No No No Yes Yes Yes No Yes Yes Yes No Yes No Good Bad Bad Bad Good Good Bad Bad Bad Good Good Bad Good No Yes Yes Yes No Yes No Yes Yes Yes Yes Yes No 25 Information Gain Example • • 5 4 I (2,3)  I (4,0) 14 14 5  I (3,2)  0.694 14 9 9 5 5 Info(D)  I (9,5)   log 2  log 2 0.940 14 14 14 14 Age  30 … >40 #yes 2 4 3 #no 3 0 2 Gain Ratio for Attribute Selection Infoage ( D)  Class P: buys_computer = es Class N: buys_computer = o • I(#yes, #no) 0.971 0 0.971 Age Income Student Credit_rating Buys_computer  30 High No Bad No  30 High No Good No … High No Bad Yes > 40 Medium No Bad Yes > 40 Low Yes Bad Yes > 40 Low Yes Good No 31...40 Low Yes Good Yes Bad No  30 Medium No  30 Low Yes Bad Yes > 40 Medium Yes Bad Yes  30 Medium Yes Good Yes 31...40 Medium No Good Yes 31...40 High Yes Bad Yes > 40 Medium No Good No • Information gain is biased towards attributes with a large number of values • Use gain ratio to normalize information gain: 5 I (2,3) ea s age  has out of 14 samples, with 2 es es a d o s. – – GainRatioA(D) = GainA(D) / SplitInfoA(D) Similar for the other terms • Hence Gain age ( D)  Info(D)  Infoage ( D)  0.246 • Similarly, Gain income ( D)  0.029 SplitInfo A( D)   Gain credit_rating ( D)  0.048 v | Dj | j 1 |D|  | Dj |   log 2   |D|  • E.g., SplitInfoincome ( D)   4 log 2 4  6 log 2 6  4 log 2 4  0.926 14 Gain student( D)  0.151 • 26 14 14 14 14 14 • GainRatioincome(D) = 0.029/0.926 = 0.031 • Attribute with maximum gain ratio is selected as splitting attribute Therefore we choose age as the splitting attribute 27 Comparing Attribute Selection Measures Gini Index • Gini index, gini(D), is defined as 28 gini( D)  1   pi2 • No clear winner (and there are many more) m i 1 – Information gain: • If data set D is split on A into v subsets D1,…, Dv, the gini index giniA(D) is defined as v |D | j gini A( D)   gini( D j ) j 1 | D | • Reduction in Impurity:  gini A( D)  gini( D)  gini A( D) • Biased towards multivalued attributes – Gain ratio: • Tends to prefer unbalanced splits where one partition is much smaller than the others – Gini index: • Biased towards multivalued attributes • Tends to favor tests that result in equal-sized partitions and purity in both partitions • Attribute that provides smallest ginisplit(D) (= largest reduction in impurity) is chosen to split the node 29 30 5 Practical Issues of Classification • • • • How Good is the Model? • Training set error: compare prediction of training record with true value Underfitting and overfitting Missing values Computational cost Expressiveness – Not a good measure for the error on unseen data. (Discussed soon.) • Test set error: for records that were not used for training, compare model prediction and true value – Use holdout data from available data set 31 32 Training versus Test Set Error • We ll eate a t ai i g dataset Output y = copy of e, except a random 25% of the records have y set to the opposite of e Five inputs, all bits, are generated in all 32 possible combinations 32 records Test Data • Generate test data using the same method: copy of e, 25% inverted; done independently from previous noise process • Some s that were corrupted in the training set will be uncorrupted in the testing set. • Some s that were uncorrupted in the training set will be corrupted in the test set. a b c d e a b c d e y y (training data) 0 0 0 0 0 0 0 0 0 0 0 0 y (test data) 0 0 0 0 0 1 0 0 0 0 0 1 0 1 0 0 0 1 0 0 0 0 0 1 0 0 1 0 0 0 1 1 1 0 0 0 1 1 1 1 0 0 1 0 0 1 0 0 1 0 0 1 1 : : : : : : : : : : : : : 1 1 1 1 1 1 1 1 1 1 1 1 1 33 Full Tree for The Training Data 34 Testing The Tree with The Test Set Root 1/4 of the tree nodes are corrupted e=0 a=0 e=1 a=1 a=0 a=1 3/4 are fine 1/16 of the test set will 1/4 of the test set records are corrupted be correctly predicted for the wrong reasons 3/16 of the test set will be wrongly predicted because the test record is corrupted 3/4 are fine 9/16 of the test predictions will be fine 3/16 of the test predictions will be wrong because the tree node is corrupted In total, we expect to be wrong on 3/8 of the test set predictions 25% of these leaf node labels will be corrupted Each leaf contains exactly one record, hence no error in predicting the training data! 35 36 6 What s This E a ple “ho Us? Suppose We Had Less Data • Discrepancy between training and test set error • But more importantly Output y = copy of e, except a random 25% of the records have y set to the opposite of e These bits are hidden 32 records – …it i di ates that the e is so ethi g e should do about it if we want to predict well on future data. a b c d e 0 0 0 0 0 y 0 0 0 0 0 1 0 0 0 0 1 0 0 0 0 0 1 1 1 0 0 1 0 0 1 : : : : : : 1 1 1 1 1 1 37 Tree Learned Without Access to The Irrelevant Bits 38 Tree Learned Without Access to The Irrelevant Bits Root e=0 Root e=1 e=0 These nodes will be unexpandable e=1 In about 12 of the 16 records in this node the output will be 0 In about 12 of the 16 records in this node the output will be 1 So this will almost certainly predict 0 So this will almost certainly predict 1 39 Tree Learned Without Access to The Irrelevant Bits Root e=0 Typical Observation Overfitting almost certainly none of the tree nodes are corrupted almost certainly all are fine 1/4 of the test set records are corrupted n/a 1/4 of the test set will be wrongly predicted because the test record is corrupted 3/4 are fine n/a 3/4 of the test predictions will be fine e=1 40 In total, we expect to be wrong on only 1/4 of the test set predictions 41 Model M overfits the training data if another model M’ exists, such that M has smaller error than M’ over the training examples, but M’ has smaller error than M over the entire distribution of instances. Underfitting: when model is too simple, both training and test errors are large 42 7 Reasons for Overfitting Avoiding Overfitting • Noise • General idea: make the tree smaller – Too losel fitti g the t ai i g data ea s the predictions reflect the noise as well – Addresses all three reasons for overfitting odel s • Prepruning: Halt tree construction early • Insufficient training data – Do not split a node if this would result in the goodness measure falling below a threshold – Difficult to choose an appropriate threshold, e.g., tree for XOR – Not enough data to enable the model to generalize beyond idiosyncrasies of the training records • Data fragmentation (special problem for trees) – Number of instances gets smaller as you traverse down the tree – Number of instances at a leaf node could be too small to make any confident decision about class • Postpruning: ‘e o e a hes f o a full g o t ee – Use a set of data different from the training data to decide when to stop pruning • Validation data: train tree on training data, prune on validation data, then test on test data 43 44 MDL-Based Pruning Intuition Minimum Description Length (MDL) X X1 X2 X3 X4 y 1 0 0 1 … … Xn 1 A? Yes No 0 B? B1 A B2 C? 1 C1 C2 0 1 B X X1 X2 X3 X4 y ? ? ? ? … … Xn ? Cost Cost(Model, Data) Cost(Model)=model size Lowest total cost • Alternative to using validation data – Motivation: data mining is about finding regular patterns in data; regularity can be used to compress the data; method that achieves greatest compression found most regularity and hence is best Cost(Data|Model)=model errors • Minimize Cost(Model,Data) = Cost(Model) + Cost(Data|Model) small – Cost is the number of bits needed for encoding. large • Cost(Data|Model) encodes the misclassification errors. • Cost(Model) uses node encoding plus splitting condition encoding. Tree size Best tree size 45 46 Handling Missing Attribute Values Distribute Instances • Missing values affect decision tree construction in three different ways: – How impurity measures are computed – How to distribute instance with missing value to child nodes – How a test instance with missing value is classified Tid Refund Marital Status Taxable Income Class 1 Yes Single 125K No 2 No Married 100K No 3 No Single 70K No 4 Yes Married 120K No 5 No Divorced 95K Yes 6 No Married No 7 Yes Divorced 220K No 8 No Single 85K Yes 9 No Married 75K No 60K Taxable Income Class 10 90K ? Single Yes 10 Refund Yes No Class=Yes 0 + 3/9 Class=Yes 2 + 6/9 Class=No 3 Class=No 4 Probability that Refund=Yes is 3/9 10 Refund Yes 47 Tid Refund Marital Status Probability that Refund=No is 6/9 No Class=Yes 0 Cheat=Yes 2 Class=No 3 Cheat=No 4 Assign record to the left child with weight = 3/9 and to the right child with weight = 6/9 48 8 Computing Impurity Measure Classify Instances Tid Refund Marital Status Taxable Income Class 1 Yes Single 125K No Split on Refund: assume records with missing values are distributed as discussed before 2 No Married 100K No 3/9 of record 10 go to Refund=Yes New record: Married Tid Refund Marital Status Taxable Income Class 11 85K No ? ? 10 3 No Single 70K No 4 Yes Married 120K No 5 No Divorced 95K Yes 6 No Married No 7 Yes Divorced 220K No 8 No Single 85K Yes 9 No Married 75K No 10 ? Single 90K Yes 60K 6/9 of record 10 go to Refund=No Entropy(Refund=Yes) – (3 / 10/3)log(3 / 10/3) = 0.469 NO Entropy(Refund=No) = -(8/3 / 20/3)log(8/3 / 20/3) Total 1 0 4 Class=Yes 6/9 1 1 2.67 Total 3.67 2 1 6.67 No MarSt Single, Divorced Married TaxInc – (4 / 20/3)log(4 / 20/3) = 0.971 Divorced 3 Refund Yes = -(1/3 / 10/3)log(1/3 / 10/3) Single Class=No < 80K NO > 80K Probability that Marital Status = Married is 3.67/6.67 10 Before Splitting: Entropy(Parent) = -0.3 log(0.3)-(0.7)log(0.7) = 0.881 Entropy(Children) = 1/3*0.469 + 2/3*0.971 = 0.804 Gain = 0.881 – 0.804 = 0.077 NO Probability that Marital Status ={Single,Divorced} is 3/6.67 49 Tree Cost Analysis 50 Tree Expressiveness • Finding an optimal decision tree is NP-complete – Optimization goal: minimize expected number of binary tests to uniquely identify any record from a given finite set • Greedy algorithm • Can represent any finite discrete-valued function – But it might not do it very efficiently • Example: parity function – Class = 1 if there is an even number of Boolean attributes with truth value = True – Class = 0 if there is an odd number of Boolean attributes with truth value = True – For accurate modeling, must have a complete tree – O(#attributes * #training_instances * log(#training_instances)) • • • • YES At each tree depth, all instances considered Assume tree depth is logarithmic (fairly balanced splits) Need to test each attribute at each node What about binary splits? – Sort data once on each attribute, use to avoid re-sorting subsets – Incrementally maintain counts for class distribution as different split points are explored • In practice, trees are considered to be fast both for training (when using the greedy algorithm) and making predictions • Not expressive enough for modeling continuous attributes – But we can still use a tree for them in practice; it just cannot accurately represent the true function 51 Classification in Large Databases Rule Extraction from a Decision Tree • One rule is created for each path from the root to a leaf • • Rules are mutually exclusive and exhaustive Example: Rule extraction from buys_computer decision-tree • Scalability: Classify data sets with millions of examples and hundreds of attributes with reasonable speed • Why use decision trees for data mining? – Precondition: conjunction of all split predicates of nodes on path – Consequent: class prediction from leaf – – – – – IF age = young AND student = no IF age = young AND student = yes IF age = mid-age IF age = old AND credit_rating = excellent IF age = young AND credit_rating = fair THEN buys_computer = no THEN buys_computer = yes THEN buys_computer = yes THEN buys_computer = yes THEN buys_computer = no – Relatively fast learning speed – Can handle all attribute types – Convertible to intelligible classification rules – Good classification accuracy, but not as good as newer methods (but tree ensembles are top!) age? <=30 31..40 student? no no yes yes yes >40 credit rating? excellent 54 fair yes 55 56 9 Scalable Tree Induction Tree Conclusions • High cost when the training data at a node does not fit in memory • Solution 1: special I/O-aware algorithm • Very popular data mining tool – Easy to understand – Easy to implement – Easy to use: little tuning, handles all attribute types and missing values – Computationally relatively cheap – Keep only class list in memory, access attribute values on disk – Maintain separate list for each attribute – Use count matrix for each attribute • Solution 2: Sampling – Common solution: train tree on a sample that fits in memory – More sophisticated versions of this idea exist, e.g., Rainforest • Overfitting problem • Focused on classification, but easy to extend to prediction (future lecture) • Build tree on sample, but do this for many bootstrap samples • Combine all into a single new tree that is guaranteed to be almost identical to the one trained from entire data set • Can be computed with two data scans 57 Classification and Prediction Overview • • • • • • • • • • 58 Theoretical Results • Trees make sense intuitively, but can we get some hard evidence and deeper understanding about their properties? • Statistical decision theory can give some answers • Need some probability concepts first Introduction Decision Trees Statistical Decision Theory Nearest Neighbor Bayesian Classification Artificial Neural Networks Support Vector Machines (SVMs) Prediction Accuracy and Error Measures Ensemble Methods 60 Random Variables 61 Working with Random Variables • E[X + Y] = E[X] + E[Y] • Var(X + Y) = Var(X) + Var(Y) + 2 Cov(X,Y) • For constants a, b • Intuitive version of the definition: – Can take on one of possibly many values, each with a certain probability – These probabilities define the probability distribution of the random variable – E.g., let X be the outcome of a coin toss, then P X= heads = . a d P X= tails = . ; dist i utio is uniform – E[aX + b] = a E[X] + b – Var(aX + b) = Var(aX) = a2 Var(X) • Iterated expectation: – E[X] = EX[ EY[Y| X] ], where EY[Y| X] = yi*Pr(Y=yi| X=x) is the expectation of Y for a given value x of X, i.e., is a function of X – In general for any function f(X,Y): EX,Y[f(X,Y)] = EX[ EY[f(X,Y)| X] ] • Consider a discrete random variable X with numeric values x1,...,xk – Expectation: E[X] =  xi*Pr(X=xi) – Variance: Var(X) = E[(X – E[X])2] = E[X2] – (E[X])2 62 63 10 What is the Optimal Model f(X)? Optimal Model f(X) (cont.) Let X denote a real - valued random input variable and Y a real - valued random output variable   The squared error of trained model f(X) is E X,Y (Y  f ( X )) 2 . f(X)  Y  E Y [Y | X ].    (Y  Y) (Y  Y)   | X   Y  f ( X )  | X   Y  f ( X )      E Y (Y  Y ) 2 | X  E Y (Y  f ( X )) 2 | X  2 E Y (Y  Y )(Y  f ( X )) | X   EY  EY 2 2 2     2      E X,Y (Y  f ( X )) 2  E X E Y (Y  Y ) 2 | X  Y  f ( X )  Consider the error for a specific value of X and let Y  E Y [Y | X ] :   Note that E X,Y (Y  f ( X )) 2  E X E Y (Y  f ( X )) 2 | X . Hence Which function f(X) will minimize the squared error? E Y (Y  f ( X )) 2 | X  E Y (Y  Y  Y  f ( X )) 2 | X  The choice of f(X) does not affect E Y (Y  Y ) 2 | X , but Y  f ( X )  is minimized for 2  Hence the squared error is minimzed by choosing f(X)  E Y [Y | X ] for every X.  2(Y  f ( X )) E Y (Y  Y ) | X  (Notice that for minimizing absolute error E X,Y | Y  f ( X ) |, one can show that the best model is 2 f(X)  median( X | Y).) (Notice : E Y (Y  Y ) | X   E Y Y | X   E Y Y | X   Y  Y  0) 64 65 Interpreting the Result • Implications for Trees To minimize mean squared error, the best prediction for input X=x is the mean of the Y-values of all training records (x(i),y(i)) with x(i)=x – E.g., assume there are training records (5,22), (5,24), (5,26), (5,28). The optimal prediction for input X=5 would be estimated as (22+24+26+28)/4 = 25. • Problem: to reliably estimate the mean of Y for a given X=x, we need sufficiently many training records with X=x. In practice, often there is only one or no training record at all for an X=x of interest. • The benefit of a good data mining technique is its ability to interpolate and extrapolate from known training records to make good predictions even for Xvalues that do not occur in the training data at all. Classification for two classes: encode as 0 and 1, use squared error as before – A tree leaf corresponds to a multi-dimensional range in the data space – Records in the same leaf are neighbors of each other – If there were many such records with X=x, we would not need a model and could just return the average Y for that X=x. • • • Since there are not enough, or none at all, training records with X=x, the output for input X=x has to be based on e o ds i the eigh o hood • Solution: estimate mean Y for input X=x from the training records in the same leaf node that contains input X=x – Classification: leaf returns majority class or class probabilities (estimated from fraction of training records in the leaf) – Prediction: leaf returns average of Y-values or fits a local model – Make sure there are enough training records in the leaf to obtain reliable estimates – Then f(X) = E[Y| X=x] = 1*Pr(Y=1| X=x) + 0*Pr(Y=0| X=x) = Pr(Y=1| X=x) Classification for k classes: can show that for 0-1 loss (error = 0 if correct class, error = 1 if wrong class predicted) the optimal choice is to return the majority class for a given input X=x – This is called the Bayes classifier. 66 Bias-Variance Tradeoff 67 Bias-Variance Tradeoff Derivation • Let s take this o e step fu the a d see if e a understand overfitting through statistical decision theory • As before, consider two random variables X and Y • From a training set D with n records, we want to construct a function f(X) that returns good approximations of Y for future inputs X – Make dependence of f on D explicit by writing f(X; D)        E X , D ,Y Y  f ( X; D)   E X ED EY Y  f ( X; D)  | X , D . Now consider the inner term : 2  2  ED EY Y  f ( X; D)  | X , D  ED EY Y  E[Y | X ] | X , D   f ( X; D)  E[Y | X ]  2   2  EY Y  E[Y | X ] | X  ED  f ( X; D)  E[Y | X ]  2 (Same derivation as before for optimal function f(X).)       (The first term does not depend on D, hence ED EY Y  E[Y | X ] | X , D  EY Y  E[Y | X ] | X .)  2  2 ED  f ( X; D)  E[Y | X ]  ED Consider the second term :  2  f ( X; D)  E [ f ( X; D)  E [ f ( X; D)]  E[Y | X ]      ED  f ( X; D)  ED [ f ( X; D)]  ED ED [ f ( X; D)]  E[Y | X ] 2 D D  2 ED  f ( X; D)  ED [ f ( X; D)   ED [ f ( X; D)]  E[Y | X ] 2  2 2  2  ED  f ( X; D)  ED [ f ( X; D)]  ED [ f ( X; D)]  E[Y | X ]  2 ED  f ( X; D)  ED [ f ( X; D) ED [ f ( X; D)]  E[Y | X ] 2   2  ED  f ( X; D)  ED [ f ( X; D)]  ED [ f ( X; D)]  E[Y | X ] (The third term is zero, because ED  f ( X; D)  ED [ f ( X; D)  ED [ f ( X; D)]  ED [ f ( X; D)]  0.) • Goal: minimize mean squared error over all X, Y, and D, i.e., EX,D,Y[ (Y - f(X; D))2 ] 2    2    E X , D ,Y Y  f ( X; D)   E X ED [ f ( X; D)]  E[Y | X ]  ED  f ( X; D)  ED [ f ( X; D)]  EY Y  E[Y | X ] | X Overall we therefore obtain : 68 2 2 2 2  69 11 Bias-Variance Tradeoff and Overfitting E D [ f ( X; D)]  E[Y | X ]2 : bias 2 E D  f ( X; D)  E D [ f ( X; D)] : variance 2 EY Y  E[Y | X ] | X : irreducible error (does not depend on f and is simply thevariance of Y given X.) • Option 1: f(X;D) = E[Y| X,D] • Option 2: f(X;D)=X (or other function independent of D) • Bias decreases as tree becomes larger – Larger tree can fit training data better • Variance increases as tree becomes larger – Bias: since ED[ E[Y| X,D] ] = E[Y| X], bias is zero – Variance: (E[Y| X,D]-ED[E[Y| X,D]])2 = (E[Y| X,D]-E[Y| X])2 can be very large since E[Y| X,D] depends heavily on D – Might overfit! – Sample variance affects predictions of larger tree more – Variance: (X-ED[X])2=(X-X)2=0 – Bias: (ED[X]-E[Y| X])2=(X-E[Y| X])2 can be large, because E[Y| X] might be completely different from X – Might underfit! • Implications for Trees • Find right tradeoff as discussed earlier – Validation data to find best pruned tree – MDL principle Find best compromise between fitting training data too closely (option 1) and completely ignoring it (option 2) 70 Classification and Prediction Overview • • • • • • • • • • 71 Lazy vs. Eager Learning • Lazy learning: Simply stores training data (or only minor processing) and waits until it is given a test record • Eager learning: Given a training set, constructs a classification model before receiving new (test) data to classify • General trend: Lazy = faster training, slower predictions • Accuracy: not clear which one is better! Introduction Decision Trees Statistical Decision Theory Nearest Neighbor Bayesian Classification Artificial Neural Networks Support Vector Machines (SVMs) Prediction Accuracy and Error Measures Ensemble Methods – Lazy method: typically driven by local decisions – Eager method: driven by global and local decisions 72 Nearest-Neighbor 73 Nearest-Neighbor Classifiers Unknown tuple • Recall our statistical decision theory analysis: Best prediction for input X=x is the mean of the Y-values of all records (x(i),y(i)) with x(i)=x (majority class for classification) • Problem was to estimate E[Y| X=x] or majority class for X=x from the training data • Solution was to approximate it – Use Y-values from training records in neighborhood around X=x • Requires: – Set of stored records – Distance metric for pairs of records ( p  q ) • Common choice: Euclidean d (p, q)  i 2 i i – Parameter k • Number of nearest neighbors to retrieve • To classify a record: – Find its k nearest neighbors – Determine output based on (distance-weighted) average of eigh o s output 74 75 12 Definition of Nearest Neighbor X X 1-Nearest Neighbor X Voronoi Diagram (a) 1-nearest neighbor (b) 2-nearest neighbor (c) 3-nearest neighbor K-nearest neighbors of a record x are data points that have the k smallest distance to x 76 Nearest Neighbor Classification 77 Effect of Changing k • Choosing the value of k: – k too small: sensitive to noise points – k too large: neighborhood may include points from other classes X Source: Hastie, Tibshirani, and Friedman. The Elements of Statistical Learning 78 Explaining the Effect of k 79 Experiment • 50 training points (x, y) • Recall the bias-variance tradeoff • Small k, i.e., predictions based on few neighbors – −2 ≤ ≤ 2, selected uniformly at random – = + �, where � is selected uniformly at random from range [-0.5, 0.5] – High variance, low bias • Test data sets: 500 points from same distribution as training data, but � = 0 • Plot 1: all (x, NN1(x)) for 5 test sets • Plot 2: all (x, AVG(NN1(x))), averaged over 200 test data set • Large k, e.g., average over entire data set – Low variance, but high bias • Need to find k that achieves best tradeoff • Can do that using validation data – Same for NN20 and NN50 80 81 13 E D [ f ( X; D)]  E[Y | X ]2 : bias 2 E D  f ( X; D)  E D [ f ( X; D)] : variance 2 EY Y  E[Y | X ] | X : irreducible error (does not depend on f and is simply thevariance of Y given X.) E D [ f ( X; D)]  E[Y | X ]2 : bias 2 E D  f ( X; D)  E D [ f ( X; D)] : variance 2 EY Y  E[Y | X ] | X : irreducible error (does not depend on f and is simply thevariance of Y given X.) E D [ f ( X; D)]  E[Y | X ]2 : bias 2 E D  f ( X; D)  E D [ f ( X; D)] : variance 2 EY Y  E[Y | X ] | X : irreducible error (does not depend on f and is simply thevariance of Y given X.) 82 84 86 E D [ f ( X; D)]  E[Y | X ]2 : bias 2 E D  f ( X; D)  E D [ f ( X; D)] : variance 2 EY Y  E[Y | X ] | X : irreducible error (does not depend on f and is simply thevariance of Y given X.) E D [ f ( X; D)]  E[Y | X ]2 : bias 2 E D  f ( X; D)  E D [ f ( X; D)] : variance 2 EY Y  E[Y | X ] | X : irreducible error (does not depend on f and is simply thevariance of Y given X.) E D [ f ( X; D)]  E[Y | X ]2 : bias 2 E D  f ( X; D)  E D [ f ( X; D)] : variance 2 EY Y  E[Y | X ] | X : irreducible error (does not depend on f and is simply thevariance of Y given X.) 83 85 87 14 Scaling Issues Other Problems • Attributes may have to be scaled to prevent distance measures from being dominated by one of the attributes • Example: • Problem with Euclidean measure: – High dimensional data: curse of dimensionality – Can produce counter-intuitive results 111111111110 100000000000 vs – Height of a person may vary from 1.5m to 1.8m – Weight of a person may vary from 90lb to 300lb – Income of a person may vary from $10K to $1M – Income difference would dominate record distance 011111111111 000000000001 d = 1.4142 d = 1.4142 – Solution: Normalize the vectors to unit length • Irrelevant attributes might dominate distance – Solution: eliminate them 88 Computational Cost 89 Classification and Prediction Overview • Brute force: O(#trainingRecords) – For each training record, compute distance to test record, keep if among top-k • Pre-compute Voronoi diagram (expensive), then search spatial index of Voronoi cells: if lucky O(log(#trainingRecords)) • Store training records in multi-dimensional search tree, e.g., R-tree: if lucky O(log(#trainingRecords)) • Bulk-compute predictions for many test records using spatial join between training and test set – Same worst-case cost as one-by-one predictions, but usually much faster in practice • • • • • • • • • • Introduction Decision Trees Statistical Decision Theory Nearest Neighbor Bayesian Classification Artificial Neural Networks Support Vector Machines (SVMs) Prediction Accuracy and Error Measures Ensemble Methods 90 Bayesian Classification 107 Bayesian Theorem: Basics • Performs probabilistic prediction, i.e., predicts class membership probabilities • Based on Bayes Theo e • Incremental training • X = a do a ia le fo data e o ds e ide e • H = hypothesis that specific record X=x belongs to class C • Goal: determine P(H| X=x) – Probability that hypothesis holds given a record x • P(H) = prior probability – Update probabilities as new training records arrive – Can combine prior knowledge with observed data – The initial probability of the hypothesis – E.g., person x will buy computer, regardless of age, income etc. • Even when Bayesian methods are computationally intractable, they can provide a standard of optimal decision making against which other methods can be measured • P(X=x) = probability that data record x is observed • P(X=x| H) = probability of observing record x, given that the hypothesis holds – E.g., given that x will buy a computer, what is the probability that x is in age group 31...40, has medium income, etc.? 108 109 15 Ba es Theo e Towards Naïve Bayes Classifier • Given data record x, the posterior probability of a hypothesis H, P(H| X=x), follows from Bayes theorem: • Suppose there are m classes C1, C2,…, Cm • Classification goal: for record x, find class Ci that has the maximum posterior probability P(Ci| X=x) • Bayes theo e : P (X  x| C )P (C ) i i P (C | X  x)  i P (X  x) P (H | X  x)  P (X  x | H )P (H ) P (X  x) • Informally: posterior = likelihood * prior / evidence • Among all candidate hypotheses H, find the maximally probably one, called maximum a posteriori (MAP) hypothesis • Note: P(X=x) is the same for all hypotheses • If all hypotheses are equally probable a priori, we only need to compare P(X=x| H) – Winning hypothesis is called the maximum likelihood (ML) hypothesis • Practical difficulties: requires initial knowledge of many probabilities and has high computational cost • Since P(X=x) is the same for all classes, only need to find maximum of P (X  x | C )P (C ) i i 110 111 Example: Computing P(X=x|Ci) and P(Ci) Computing P(X=x|Ci) and P(Ci) • Estimate P(Ci) by counting the frequency of class Ci in the training data • Can we do the same for P(X=x|Ci)? – Need very large set of training data – Have |X1|*|X2|*…*|Xd|*m different combinations of possible values for X and Ci – Need to see every instance x many times to obtain reliable estimates • Solution: decompose into lower-dimensional problems • • • P(buys_computer = yes) = 9/14 P(buys_computer = no) = 5/14 P(age>40, income=low, student=no, credit_rating=bad| buys_computer=yes) = 0 ? Age Income Student Credit_rating Buys_computer High No Bad No  30 High No Good No  30 … High No Bad Yes > 40 Medium No Bad Yes > 40 Low Yes Bad Yes Low Yes Good No > 40 31...40 Low Yes Good Yes  30 Medium No Bad No  30 Low Yes Bad Yes > 40 Medium Yes Bad Yes Good Yes  30 Medium Yes 31...40 Medium No Good Yes 31...40 High Yes Bad Yes > 40 Medium No Good No 112 Conditional Independence 113 Derivation of Naïve Bayes Classifier • Simplifying assumption: all input attributes conditionally independent, given class • X, Y, Z random variables • X is conditionally independent of Y, given Z, if P(X| Y,Z) = P(X| Z) P ( X  ( x1 ,, xd ) | Ci )   P ( Xk  xk | Ci )  P ( X1  x1 | Ci )  P ( X2  x2 | Ci )  P ( Xd  xd | Ci ) d k 1 – Equivalent to: P(X,Y| Z) = P(X| Z) * P(Y| Z) • Each P(Xk=xk| Ci) can be estimated robustly • Example: people with longer arms read better – If Xk is categorical attribute – Confounding factor: age • Young child has shorter arms and lacks reading skills of adult – If age is fixed, observed relationship between arm length and reading skills disappears • P(Xk=xk| Ci) = #records in Ci that have value xk for Xk, divided by #records of class Ci in training data set – If Xk is continuous, we could discretize it • Problem: interval selection – Too many intervals: too few training cases per interval – Too few intervals: limited choices for decision boundary 114 115 16 Estimating P(Xk=xk| Ci) for Continuous Attributes without Discretization • P(Xk=xk| Ci) computed based on Gaussian distribution with mean μ and standard deviation σ: ( x  ) 2  1 2 g ( x,  ,  )  e 2 2  as P( Xk  xk | C i)  g ( xk , k,Ci ,  k,Ci ) Naïve Bayes Example • Classes: – C1:buys_computer = yes – C2:buys_computer = no • Data sample x • Estimate k,Ci from sample mean of attribute Xk for all training records of class Ci • Estimate k,Ci similarly from sample – – – – age  30, income = medium, student = yes, and credit_rating = bad Age Income Student Credit_rating Buys_computer No Bad No  30 High No Good No  30 High … High No Bad Yes Bad Yes > 40 Medium No > 40 Low Yes Bad Yes > 40 Low Yes Good No 31...40 Low Yes Good Yes  30 Medium No Bad No  30 Low Yes Bad Yes > 40 Medium Yes Bad Yes Good Yes  30 Medium Yes 31...40 Medium No Good Yes 31...40 High Yes Bad Yes Good No > 40 Medium No 116 Naïve Bayesian Computation • Compute P(Ci) for each class: • Compute P(Xk=xk| Ci) for each class – – – – – – – – – – P(buys_computer = P(buys_computer = • Compute final result P(X=x| Ci) * P(Ci) • Therefore we predict buys_computer = es fo input x = age =  , i o e = ediu , stude t = – – • Naïve Bayesian prediction requires each conditional probability to be non-zero (why?) P ( X  ( x1 ,, xd ) | Ci )   P ( Xk  xk | Ci )  P ( X1  x1 | Ci )  P ( X2  x2 | Ci )  P ( Xd  xd | Ci ) P age =  | buys_computer = es = / = . P age =  | buys_computer = o = / = . P i o e = ediu | buys_computer = es = / = . P i o e = ediu | buys_computer = o = / = . P stude t = es | buys_computer = es = / = . P stude t = es | buys_computer = o = / = . P(credit_rating = ad | buys_computer = es = / = . P(credit_rating = ad | buys_computer = o = / = . Compute P(X=x| Ci) using the Naive Bayes assumption – – Zero-Probability Problem es = / = . o = / = . • P(30, medium, yes, fair |buys_computer = es = . * . * . * . P(30, medium, yes, fair | buys_computer = o = . * . * . * . = . P(X=x | buys_computer = P(X=x | buys_computer = 117 d k 1 • Example: 1000 records for buys_computer=yes with income=low (0), income= medium (990), and income = high (10) – For input with income=low, conditional probability is zero • Use Laplacian correction (or Laplace estimator) by adding 1 dummy record to each income level = . • Prob(income = low) = 1/1003 • Prob(income = medium) = 991/1003 • Prob(income = high) = 11/1003 es * P buys_computer = es = . o * P buys_computer = o = . es , credit_rating = – Co e ted p o a ilit esti ates lose to thei counterparts, but none is zero ad u o e ted 118 119 Probabilities Naïve Bayesian Classifier: Comments • Easy to implement • Good results obtained in many cases – Robust to isolated noise points – Handles missing values by ignoring the instance during probability estimate calculations – Robust to irrelevant attributes • Disadvantages – Assumption: class conditional independence, therefore loss of accuracy – Practically, dependencies exist among variables • Summary of elementary probability facts we have used already and/or will need soon • Let X be a random variable as usual • Let A be some predicate over its possible values – A is true for some values of X, false for others – E.g., X is out o e of th o of a die, A ould e is g eate tha alue • P(A) is the fraction of possible worlds in which A is true • How to deal with these dependencies? – P(die value is greater than 4) = 2 / 6 = 1/3 120 121 17 Axioms • • • • Theorems from the Axioms 0  P(A)  1 P(True) = 1 P(False) = 0 P(A  B) = P(A) + P(B) - P(A  B) • 0  P(A)  1, P(True) = 1, P(False) = 0 • P(A  B) = P(A) + P(B) - P(A  B) • From these we can prove: – P(not A) = P(~A) = 1 - P(A) – P(A) = P(A  B) + P(A  ~B) 122 Conditional Probability Definition of Conditional Probability P(A  B) P(A| B) = -----------P(B) • P(A|B) = Fraction of worlds in which B is true that also have A true H = “Have a headache” F = “Coming down with Flu” P(H) = 1/10 P(F) = 1/40 P(H|F) = 1/2 F H 123 Corollary: the Chain Rule P(A  B) = P(A| B) P(B) “Headaches are rare and flu is rarer, but if you’re coming down with flu there’s a 5050 chance you’ll have a headache.” 124 125 Easy Fact about Multivalued Random Variables Multivalued Random Variables • Suppose X can take on more than 2 values • X is a random variable with arity k if it can take on exactly one value out of {v1, v2,…, vk} • Thus P ( X  vi  X  v j )  0 if i  j • Using the axioms of probability – 0  P(A)  1, P(True) = 1, P(False) = 0 – P(A  B) = P(A) + P(B) - P(A  B) • And assuming that X obeys P ( X  vi  X  v j )  0 if i  j P ( X  v1  X  v2  ...  X  vk )  1 i • We can prove that P ( X  v1  X  v2  ...  X  vi )   P ( X  v j ) P ( X  v1  X  v2  ...  X  vk )  1 • And therefore: 126  P( X  v )  1 k j 1 j 1 j 127 18 Useful Easy-to-Prove Facts The Joint Distribution Recipe for making a joint distribution of d variables: P ( A| B)P (~ A| B)  1  P( X  v k j 1 j Example: Boolean variables A, B, C | B)  1 128 The Joint Distribution Recipe for making a joint distribution of d variables: 1. Make a truth table listing all combinations of values of your variables (has 2d rows for d Boolean variables). The Joint Distribution Example: Boolean variables A, B, C A B C 0 0 0 0 0 1 0 1 0 0 1 1 1 0 0 1 0 1 1 1 0 1 1 1 129 Recipe for making a joint distribution of d variables: 1. Make a truth table listing all combinations of values of your variables (has 2d rows for d Boolean variables). 2. For each combination of values, say how probable it is. Example: Boolean variables A, B, C A B C Prob 0 0 0 0.30 0 0 1 0.05 0 1 0 0.10 0 1 1 0.05 1 0 0 0.05 1 0 1 0.10 1 1 0 0.25 1 1 1 0.10 130 The Joint Distribution Recipe for making a joint distribution of d variables: 1. Make a truth table listing all combinations of values of your variables (has 2d rows for d Boolean variables). 2. For each combination of values, say how probable it is. 3. If you subscribe to the axioms of probability, those numbers must sum to 1. B C Prob 0 0 0 0.30 0 0 1 0.05 0 1 0 0.10 0 1 1 0.05 1 0 0 0.05 1 0 1 0.10 1 1 0 0.25 1 1 1 0.10 A 0.05 0.25 0.30 Using the Joint Dist. Example: Boolean variables A, B, C A B 0.10 Once you have the JD you can ask for the probability of any logical expression involving your attribute 0.05 0.10 0.05 131 P (E)   P (row) rows matching E C 0.10 132 133 19 Using the Joint Dist. Using the Joint Dist. P(Poor  Male) = 0.4654 P (E)   P (row) P (E)  P(Poor) = 0.7604 rows matching E  P (row) rows matching E 134 Inference with the Joint Dist. P ( E1 | E2 )  135 Inference with the Joint Dist. P ( E1  E2 )  P ( E2 )  P (row )  P (row ) P ( E1 | E2 )  rows matching E1 and E2 P ( E1  E2 )  P ( E2 ) rows matching E2  P (row )  P (row ) rows matching E1 and E2 rows matching E2 P(Male | Poor) = 0.4654 / 0.7604 = 0.612 136 Joint Distributions • Good news: Once you have a joint distribution, you can answer important questions that involve uncertainty. 137 What Would Help? • Bad news: Impossible to create joint distribution for more than about ten attributes because there are so many numbers needed when you build it. • Full independence – P(gender=g  hours_worked=h  wealth=w) = P(gender=g) * P(hours_worked=h) * P(wealth=w) – Can reconstruct full joint distribution from a few marginals • Full conditional independence given class value – Naïve Bayes • What about something between Naïve Bayes and general joint distribution? 138 139 20 Bayesian Belief Networks Bayesian Network Properties • Subset of the variables conditionally independent • Graphical model of causal relationships – Represents dependency among the variables – Gives a specification of joint probability distribution  Nodes: random variables  Links: dependency Y  X and Y are the parents of Z, and Y is Y X • Each variable is conditionally independent of its non-descendents in the graph, given its parents • Naïve Bayes as a Bayesian network: the parent of P Z P  Given Y, Z and P are independent X1  Has no loops or cycles X2 Xn 140 141 General Properties Structural Property • P(X1,X2,X3)=P(X1|X2,X3)P(X2|X3)P(X3) • P(X1,X2,X3)=P(X3|X1,X2)P(X2|X1)P(X1) • Network does not necessarily reflect causality X1 X2 X1 X3 • Missing links simplify computation of P � , � , … , �� • General: ��= P �� |��− , ��− , … , � – Fully connected: link between every pair of nodes • Given network: � �= P �� |parents �� – Some links are missing – The terms P �� |parents �� are given as conditional probability tables (CPT) in the network • “pa se et o k allo s ette esti atio of CPT s (fewer combinations of parent values, hence more reliable to estimate from limited data) and faster computation X2 X3 142 143 Small Example Computing P(S|J) • S: Student studies a lot for 6220 • L: Student learns a lot and gets a good grade • J: Student gets a great job P(S) = 0.4 S L P(J|L) = 0.8 P(J|~L) = 0.3 P(L|S) = 0.9 P(L|~S) = 0.2 • • Probability that a student who got a great job was doing her homework P(S | J) = P(S, J) / P(J) • • P(S, J) = P(S, J, L) + P(S, J, ~L) P(J) = P(J, S, L) + P(J, S, ~L) + P(J, ~S, L) + P(J, ~S, ~L) • • • • P(J, L, S) = P(J | L, S) * P(L, S) = P(J | L) * P(L | S) * P(S) = 0.8*0.9*0.4 P(J, ~L, S) = P(J | ~L, S) * P(~L, S) = P(J | ~L) * P(~L | S) * P(S) = 0.3*(1-0.9)*0.4 P(J, L, ~S) = P(J | L, ~S) * P(L, ~S) = P(J | L) * P(L | ~S) * P(~S) = 0.8*0.2*(1-0.4) P(J, ~L, ~S) = P(J | ~L, ~S) * P(~L, ~S) = P(J | ~L) * P(~L | ~S) * P(~S) = 0.3*(1-0.2)*(10.4) • • Putting this all together, we obtain: P(H | J) = (0.8*0.9*0.4 + 0.3*0.1*0.4) / (0.8*0.9*0.4 + 0.3*0.1*0.4 + 0.8*0.2*0.6 + 0.3*0.8*0.6) = 0.3 / 0.54 = 0.56 J 144 145 21 More Complex Example Computing with Bayes Net P(S)=0.3 T: The lecture started on time L: The lecturer arrives late R: The lecture concerns data mining M: The lecturer is Mike S: It is snowing M P(LM, S)=0.05 P(LM, ~S)=0.1 P(L~M, S)=0.1 P(L~M, ~S)=0.2 R M S P(RM)=0.3 P(R~M)=0.6 L L R P(TL)=0.3 P(T~L)=0.8 T ? S P(M)=0.6 T: The lecture started on time L: The lecturer arrives late R: The lecture concerns data mining M: The lecturer is Mike S: It is snowing P(T, ~R, L, ~M, S) = P(T | L)  P(~R | ~M)  P(L | ~M, S)  P(~M)  P(S) T 146 147 Computing with Bayes Net P(S)=0.3 M S P(LM, S)=0.05 P(LM, ~S)=0.1 P(L~M, S)=0.1 P(L~M, ~S)=0.2 P(RM)=0.3 P(R~M)=0.6 P(TL)=0.3 P(T~L)=0.8 P(R | T, ~S) = P(R, T, ~S) / P(T, ~S) • Can predict the probability for any attribute, given any subset of the other attributes P(M)=0.6 L T Inference with Bayesian Networks – P(M | L, R), P(T | S, ~M, R) and so on R T: The lecture started on time L: The lecturer arrives late R: The lecture concerns data mining M: The lecturer is Mike S: It is snowing • Easy case: P(Xi | Xj1, Xj2,…, Xjk) where parents(Xi){Xj1, Xj2,…, Xjk} – Can read answer directly from Xi s CPT • What if values are not given for all parents of Xi? – Exact inference of probabilities in general for an arbitrary Bayesian network is NP-hard – Solutions: probabilistic inference, trade precision for efficiency P(R, T, ~S) = P(L, M, R, T, ~S) + P(~L, M, R, T, ~S) + P(L, ~M, R, T, ~S) + P(~L, ~M, R, T, ~S) Compute P(T, ~S) similarly. Problem: There are now 8 such terms to be computed. 148 Training Bayesian Networks 149 Classification and Prediction Overview • Several scenarios: – Network structure known, all variables observable: learn only the CPTs – Network structure known, some hidden variables: gradient descent (greedy hill-climbing) method, analogous to neural network learning – Network structure unknown, all variables observable: search through the model space to reconstruct network topology – Unknown structure, all hidden variables: No good algorithms known for this purpose • Ref.: D. Heckerman: Bayesian networks for data mining 150 • • • • • • • • • • Introduction Decision Trees Statistical Decision Theory Nearest Neighbor Bayesian Classification Artificial Neural Networks Support Vector Machines (SVMs) Prediction Accuracy and Error Measures Ensemble Methods 152 22 Basic Building Block: Perceptron For Example d   f (x)  sign  b   wi xi  i 1   Called the bias x1 w1 x2 w2 xd wd  Perceptron Decision Hyperplane {(x1, x2, , …} Input: x2 Output: classification function f(x) f(x) > 0: return +1 f(x ≤ : etu = -1 b+w1x1+w2x2 = 0 +b Decision hyperplane: b+w∙x = 0 f Output y  w x  b Note: b+w∙x > 0, if and only if d Input Weight vector x vector w Weighted sum i 1 Activation function x1 i i b represents a threshold for when the perceptron fi es . 153 154 Representing Boolean Functions Perceptron Training Rule • AND with two-input perceptron – b=-0.8, w1=w2=0.5 • OR with two-input perceptron – b=-0.3, w1=w2=0.5 • m-of-n function: true if at least m out of n inputs are true – All input weights 0.5, threshold weight b is set according to m, n • Goal: correct +1/-1 output for each training record • Start with random weights, constant  (learning rate) • While some training records are still incorrectly classified do – For each training record (x, y) • Let fold(x) be the output of the current perceptron for x • Set b:= b + b, where b = ( y - fold(x) ) • For all i, set wi := wi + wi, where wi = ( y - fold(x))xi • Converges to correct decision boundary, if the classes are linearly separable and a small enough  is used • Can also represent NAND, NOR • What about XOR? 155 156 Gradient Descent Gradient Descent Rule • If training records are not linearly separable, find best fit approximation • Find weight vector that minimizes E(b,w) by altering it in direction of steepest descent – Set (b,w) := (b,w) + (b,w), where (b,w) = - E(b,w) – Gradient descent to search the space of possible weight vectors – Basis for Backpropagation algorithm • -E(b,w)=[ E/b, E/w1,…, E/wn ] is the gradient, hence • Consider un-thresholded perceptron (no sign function applied), i.e., u(x) = b + w∙x • Measure training error by squared error 1 2 E(b, w)    y  u(x) 2 ( x, y)D – D = training data   E  b       y  u(x)  b  ( x, y)D  E wi : wi    wi    y  u(x) ( xi ) wi ( x , y)D b : b   • Start with random weights, iterate until convergence – Will converge to global minimum if  is small enough E(w1,w2) 100 90 80 70 60 50 40 30 20 10 0 4 2 -4 0 -2 w1 157 0 -2 2 4 w2 -4 158 23 Gradient Descent Summary Multilayer Feedforward Networks • Epoch updating (batch mode) • Use another perceptron to combine output of lower layer – Compute gradient over entire training set – Changes model once per scan of entire training set Output layer – What about linear units only? Can only construct linear functions! – Need nonlinear component • Case updating (incremental mode, stochastic gradient descent) • sign function: not differentiable (gradient descent!) • Use sigmoid: (x)=1/(1+e-x) – Compute gradient for a single training record – Changes model after every single training record immediately • Case updating can approximate epoch updating arbitrarily close if  is small enough • What is the difference between perceptron training rule and case updating for gradient descent? – Error computation on thresholded vs. unthresholded function 1 Hidden layer 1/(1+exp(-x)) 0.9 0.8 0.7 0.6 Perceptron function: 0.5 y 0.4 0.3 0.2 0.1 1 1  e b wx Input layer 0 159 1-Hidden Layer ANN Example   NINS v1  g  b1   w1k xk  k 1   w11 w21 x1 w31 w12 w22 x2 w32 N INS   v2  g  b2   w2 k xk  k 1   N INS   v3  g  b3   w3k xk  k 1   -2 0 2 4 160 Making Predictions w1 w2 -4 N HID   Out  g  B   Wk vk  k 1   w3 g is usually the sigmoid function • Input record fed simultaneously into the units of the input layer • Then weighted and fed simultaneously to a hidden layer • Weighted outputs of the last hidden layer are the input to the units in the output layer, which emits the network's prediction • The network is feed-forward – None of the weights cycles back to an input unit or to an output unit of a previous layer • Statistical point of view: neural networks perform nonlinear regression 161 Backpropagation Algorithm 162 Overfitting • Earlier discussion: gradient descent for a single perceptron using a simple un-thresholded function • If sigmoid (or other differentiable) function is applied to weighted sum, use complete function for gradient descent • Multiple perceptrons: optimize over all weights of all perceptrons – Problems: huge search space, local minima • Backpropagation • When do we stop updating the weights? • Overfitting tends to happen in later iterations – Weights initially small random values – Weights all similar => smooth decision surface – Surface complexity increases as weights diverge • Preventing overfitting – Initialize all weights with small random values – Iterate many times – Weight decay: decrease each weight by small factor during each iteration, or – Use validation data to decide when to stop iterating • Compute gradient, starting at output and working back – Error of hidden unit h: how do we get the true output value? Use weighted sum of errors of each unit influenced by h • Update all weights in the network 163 164 24 Backpropagation Remarks Neural Network Decision Boundary • Computational cost – Each iteration costs O(|D|*|w|), with |D| training records and |w| weights – Number of iterations can be exponential in n, the number of inputs (in practice often tens of thousands) • Local minima can trap the gradient descent algorithm: convergence guaranteed to local minimum, not global • Backpropagation highly effective in practice – Many variants to deal with local minima issue, use of case updating Source: Hastie, Tibshirani, and Friedman. The Elements of Statistical Learning 165 166 Defining a Network Representational Power • Boolean functions 1. Decide network topology – #input units, #hidden layers, #units per hidden layer, #output units (one output unit per class for problems with >2 classes) – Each can be represented by a 2-layer network – Number of hidden units can grow exponentially with number of inputs 2. Normalize input values for each attribute to [0.0, 1.0] – Nominal/ordinal attributes: one input unit per domain value • Create hidden unit for each input record • Set its weights to activate only for that input • Implement output unit as OR gate that only activates for desired output patterns • For attribute grade with values A, B, C, have 3 inputs that are set to 1,0,0 for grade A, to 0,1,0 for grade B, and 0,0,1 for C • Why not map it to a single input with domain [0.0, 1.0]? 3. Choose learning rate , e.g., 1 / (#training iterations) • Continuous functions – Too small: takes too long to converge – Too large: might never converge (oversteps minimum) 4. Bad results on test data? Change network topology, initial weights, or learning rate; try again. – Every bounded continuous function can be approximated arbitrarily close by a 2-layer network • Any function can be approximated arbitrarily close by a 3-layer network 167 Neural Network as a Classifier Classification and Prediction Overview • • • • • • • • • • • Weaknesses – Long training time – Many non-trivial parameters, e.g., network topology – Poor interpretability: What is the meaning behind learned weights and hidden units? • Note: hidden units are alternative representation of input values, capturing their relevant features • Strengths – – – – 168 High tolerance to noisy data Well-suited for continuous-valued inputs and outputs Successful on a wide array of real-world data Techniques exist for extraction of rules from neural networks 169 Introduction Decision Trees Statistical Decision Theory Nearest Neighbor Bayesian Classification Artificial Neural Networks Support Vector Machines (SVMs) Prediction Accuracy and Error Measures Ensemble Methods 171 25 SVM—Support Vector Machines SVM—History and Applications • Newer and very popular classification method • Uses a nonlinear mapping to transform the original training data into a higher dimension • Searches for the optimal separating hyperplane i.e., de isio ou da i the new dimension • SVM finds this hyperplane using support e to s esse tial t ai i g e o ds a d margins (defined by the support vectors) • Vapnik and colleagues (1992) – Groundwork from Vapnik & Chervonenkis statisti al learning theory in 1960s • Training can be slow but accuracy is high – Ability to model complex nonlinear decision boundaries (margin maximization) • Used both for classification and prediction • Applications: handwritten digit recognition, object recognition, speaker identification, benchmarking time-series prediction tests 172 173 Linear Classifiers denotes +1 Linear Classifiers f(x,w,b) = sign(wx + b) denotes -1 denotes +1 f(x,w,b) = sign(wx + b) denotes -1 How would you classify this data? How would you classify this data? 174 175 Linear Classifiers denotes +1 Linear Classifiers f(x,w,b) = sign(wx + b) denotes -1 denotes +1 f(x,w,b) = sign(wx + b) denotes -1 How would you classify this data? 176 How would you classify this data? 177 26 Linear Classifiers denotes +1 Classifier Margin f(x,w,b) = sign(wx + b) denotes -1 f(x,w,b) = sign(wx + b) denotes +1 Define the margin of a linear classifier as the width that the boundary could be increased by before hitting a data record. denotes -1 Any of these would be fine.. ..but which is best? 178 179 Maximum Margin denotes +1 denotes -1 Maximum Margin f(x,w,b) = sign(wx + b) Find the maximum margin linear classifier. This is the simplest kind of SVM, called linear SVM or LSVM. f(x,w,b) = sign(wx + b) denotes +1 denotes -1 Support Vectors are those datapoints that the margin pushes up against 180 Why Maximum Margin? 181 Specifying a Line and Margin Plus-Plane • If we made a small error in the location of the boundary, this gives us the least chance of causing a misclassification. • Model is immune to removal of any nonsupport-vector data records. • There is some theory (using VC dimension) that is related to (but not the same as) the proposition that this is a good thing. • Empirically it works very well. 182 Classifier Boundary Minus-Plane • Plus-plane = { x : wx + b = +1 } • Minus-plane = { x : wx + b = -1 } Classify as w x + b  1 +1 if -1 if wx + b  -1 what if -1 < wx + b < 1 ? 183 27 Computing Margin Width Computing Margin Width M = Margin Width M = Margin Width x+ x- • Plus-plane = { x : wx + b = +1 } • Minus-plane = { x : wx + b = -1 } • Goal: compute M in terms of w and b – Note: vector w is perpendicular to plus-plane • Consider two vectors u and v on plus-plane and show that w(u-v)=0 • Hence it is also perpendicular to the minus-plane • Choose arbitrary point x- on minus-plane • Let x+ be the point in plus-plane closest to x• Since vector w is perpendicular to these planes, it holds that x+ = x- + w, for some value of  184 Putting It All Together 185 Finding the Maximum Margin • We have so far: • How do we find w and b such that the margin is maximized and all training records are in the correct zone for their class? • Solution: Quadratic Programming (QP) • QP is a well-studied class of optimization algorithms to maximize a quadratic function of some real-valued variables subject to linear constraints. – wx+ + b = +1 and wx- + b = -1 – x+ = x- + w – |x+- x-| = M • Derivation: – w(x- + w) + b = +1, hence wx- + b + ww = 1 – This implies ww = 2, i.e.,  = 2 / ww – Since M = |x+- x-| = |w| =  |w| = (ww)0.5 – We obtain M = 2 (ww)0.5/ ww = 2 / (ww)0.5 – There exist algorithms for finding such constrained quadratic optima efficiently and reliably. 186 Quadratic Programming Find uT Ru arg max c  d u  2 u Subject to T a11u1  a12u2  ...  a1mum  b1 a 21u1  a 22u2  ...  a 2 mum  b2 : a n1u1  a n 2u2  ...  a nmum  bn And subject to 187 What Are the SVM Constraints? Quadratic criterion n additional linear inequality constraints a ( n 1)1u1  a ( n 1) 2u2  ...  a ( n 1) mum  b( n 1) a ( n  2)1u1  a ( n  2) 2u2  ...  a ( n  2) mum  b( n  2) : a ( n  e )1u1  a ( n  e ) 2u2  ...  a ( n  e ) mum  b( n  e ) M 2 ww • What is the quadratic optimization criterion? • Consider n training records (x(k), y(k)), where y(k) = +/- 1 • How many constraints will we have? • What should they be? e additional linear equality constraints 188 189 28 Problem: Classes Not Linearly Separable What Are the SVM Constraints? M 2 ww • What is the quadratic optimization criterion? – Minimize ww • Consider n training records (x(k), y(k)), where y(k) = +/- 1 • How many constraints will we have? n. • What should they be? For each 1  k  n: wx(k) + b  1, if y(k)=1 wx(k) + b  -1, if y(k)=-1 • Inequalities for training records are not satisfiable by any w and b denotes +1 denotes -1 190 191 Solution 1? denotes +1 denotes -1 Solution 2? • Find minimum ww, while also minimizing number of training set errors • Minimize ww + C(#trainSetErrors) denotes +1 denotes -1 – C is a tradeoff parameter • Problems: – Cannot be expressed as QP, hence finding solution might be slow – Does not distinguish between disastrous errors and near misses – Not a well-defined optimization problem (cannot optimize two things at the same time) 192 Solution 3 denotes +1 denotes -1 193 What Are the SVM Constraints? • Minimize ww + C(distance of error records to their correct place) • This works! • But still need to do something about the unsatisfiable set of inequalities 194 2 11 7 M 2 ww • What is the quadratic optimization criterion? – Minimize n 1 w  w  C  εk 2 k 1 • Consider n training records (x(k), y(k)), where y(k) = +/- 1 • How many constraints will we have? n. • What should they be? For each 1  k  n: wx(k)+b  1 - k, if y(k)=1 wx(k)+b  -1+k, if y(k)=-1 k  0 195 29 Facts About the New Problem Formulation Effect of Parameter C • Original QP formulation had d+1 variables – w1, w2,..., wd and b • New QP formulation has d+1+n variables – w1, w2,..., wd and b – 1, 2,..., n • C is a new parameter that needs to be set for the SVM – Controls tradeoff between paying attention to margin size versus misclassifications Source: Hastie, Tibshirani, and Friedman. The Elements of Statistical Learning 196 A E ui ale t QP The Dual α n Maximize k 1 k Subject to these constraints:  Important Facts 1  αkαl  y(k)  y(l )  x(k)  x(l ) 2 k 1 l 1 n n α k : 0  αk  C w   αk  y(k)  x(k) n k 1 k y(k)  0 Then define:   1  x( k )  w  b  AVG  k:0 k C y( k )   • Dual formulation of QP can be optimized more quickly, but result is equivalent • Data records with k > 0 are the support vectors – Those with 0 < k < C lie on the plus- or minus-plane – Those with k = C are on the wrong side of the classifier boundary (have k > 0) • Computation for w and b only depends on those records with k > 0, i.e., the support vectors • Alternative QP has another major advantage, as we will see now... n k 1 197 Then classify with: f(x,w,b) = sign(wx + b) 198 Easy To Separate 199 Easy To Separate What would SVMs do with this data? Not a big surprise Positive “plane” 200 Negative “plane” 201 30 Harder To Separate Harder To Separate X’ (= X2) What can be done about this? Non-linear basis functions: Original data: (X, Y) Transformed: (X, X2, Y) X Think of X2 as a new attribute, e.g., X’ 202 203 Co espo di g Pla es i O igi al Space Now Separation Is Easy Again X’ (= X2) Region above plus-”plane” X Region below minus-”plane” 204 Common SVM Basis Functions • Polynomial of attributes X1,..., Xd of certain max degree, e.g., X42 • Radial basis function – Symmetric around center, i.e., KernelFunction(|X - c| / kernelWidth) • Sigmoid function of X, e.g., hyperbolic tangent • Let (x) be the transformed input record – Previous example: ( (x) ) = (x, x2) 206 205 1     2 x1    2 x2    :     x 2 d     x12   2 x 2     :   2  xd  Φ(x)   2 x1 x2     2 x1 x3    :    2 x1 xd    x x 2 2 3     :    2 x1 xd    :  2 x x  d 1 d   Constant Term Linear Terms Pure Quadratic Terms Quadratic Basis Functions Number of terms (assuming d input attributes): (d+2)-choose-2 = (d+2)(d+1)/2  d2/2 Quadratic Cross-Terms Why did we choose this specific transformation? 207 31 Dual QP With Basis Functions α n Maximize k 1 k  Subject to these constraints: 1  αkαl  y(k)  y(l )  Φx(k) Φx(l ) 2 k 1 l 1 n n α k : 0  αk  C w   αk  y(k)  Φx(k)  n k 1 k y(k)  0 Then define: n k 1  1   Φx(k)   w  b  AVG  k:0 k C y( k )   Then classify with: f(x,w,b) = sign(w(x) + b) Computation Challenge • Input vector x has d components (its d attribute values) • The transformed input vector (x) has d2/2 components • Hence computing (x(k))(x(l)) now costs order d2/2 instead of order d operations (additions, multiplications) • ...or is there a better way to do this? – Take advantage of properties of certain transformations 208 1 1         2b1  2 a1       2b2  2a 2     : :        2b  a 2 d d     2 2  a1   b1      2 2 a b 2 2         : :     2 2  ad   bd   Φ(a)  Φ(b)    2b1b2  2 a1a 2      2a1a 3   2b1b3      : :      2a1a d   2b1bd      a a 2 b b 2 2 3   2 3       : :      2a1a d   2b1bd      : :  2a a   2b b  d 1 d   d 1 d   Quadratic Dot Products 1  2a b + d i 1 209 i i Quadratic Dot Products Now consider another function of a and b: + a b (a b  1) 2 d 2 2 i i i 1 Φ(a)  Φ(b)  1  2 a i bi   a i2bi2    2a i a j bi b j +   2a a b b d d d i 1 i 1 d d i 1 j i 1 i j i d  d     a i bi   2 a i bi  1 i 1  i 1  2   a i bi a j b j  2 a i bi  1 d d d i 1 j 1 d i 1 j i 1  (a  b) 2  2a  b  1 i 1   (a i bi ) 2  2  a i bi a j b j  2 a i bi  1 j d i 1 d d i 1 j i 1 210 Quadratic Dot Products d i 1 211 Any Other Computation Problems? • The results of (a)(b) and of (ab+1)2 are identical • Computing (a)(b) costs about d2/2, while computing (ab+1)2 costs only about d+2 operations • This means that we can work in the high-dimensional space (d2/2 dimensions) where the training records are more easily separable, but pay about the same cost as working in the original space (d dimensions) • Savings are even greater when dealing with higherdegree polynomials, i.e., degree q>2, that can be computed as (ab+1)q n  1  w   αk  y(k)  Φx(k)  b  AVG   Φx(k)   w     k : 0 C k ( ) y k k 1   • What about computing w? w  Φ(x)   αk  y(k)  Φx(k)   Φ(x) – Finally need f(x,w,b) = sign(w(x) + b): n k 1 – Can be computed using the same trick as before • Can apply the same trick again to b, because Φx(k)   w   α j  y( j )  Φx(k)   Φx( j )  n j 1 212 213 32 SVM Kernel Functions Overfitting • For which transformations, called kernels, does the same trick work? • Polynomial: K(a,b)=(a  b +1)q • Radial-Basis-style (RBF): q, , , and  are  (a  b) 2   K(a, b)  exp   2 2   – Neural-net-style sigmoidal: magic parameters that must be chosen by a model selection method. K(a, b)  tanh(  a  b   ) • With the right kernel function, computation in high dimensional transformed space is no problem • But what about overfitting? There seem to be so many parameters... • Usually not a problem, due to maximum margin approach – Only the support vectors determine the model, hence SVM complexity depends on number of support vectors, not dimensions (still, in higher dimensions there might be more support vectors) – Minimizing ww discourages extremely large weights, which smoothes the function (recall weight decay for neural networks!) 214 215 Different Kernels Multi-Class Classification • SVMs can only handle two-class outputs (i.e. a categorical output variable with arity 2). • With output arity N, lea N “VM s – – – – Source: Hastie, Tibshirani, and Friedman. The Elements of Statistical Learning “VM lea s Output== vs Output != “VM lea s Output== vs Output != : “VM N lea s Output==N vs Output != N • Predict with each SVM and find out which one puts the prediction the furthest into the positive region. 216 217 Why Is SVM Effective on High Dimensional Data? SVM vs. Neural Network • Complexity of trained classifier is characterized by the number of support vectors, not dimensionality of the data • If all other training records are removed and training is repeated, the same separating hyperplane would be found • The number of support vectors can be used to compute an upper bound on the expected error rate of the SVM, which is independent of data dimensionality • Thus, an SVM with a small number of support vectors can have good generalization, even when the dimensionality of the data is high 218 • SVM – Relatively new concept – Deterministic algorithm – Nice Generalization properties – Hard to train – learned in batch mode using quadratic programming techniques – Using kernels can learn very complex functions • Neural Network – Relatively old – Nondeterministic algorithm – Generalizes well but does t ha e st o g mathematical foundation – Can easily be learned in incremental fashion – To learn complex functions—use multilayer perceptron (not that trivial) 219 33 What Is Prediction? Classification and Prediction Overview • • • • • • • • • • • Essentially the same as classification, but output is continuous, not discrete Introduction Decision Trees Statistical Decision Theory Nearest Neighbor Bayesian Classification Artificial Neural Networks Support Vector Machines (SVMs) Prediction Accuracy and Error Measures Ensemble Methods – Construct a model, then use model to predict continuous output value for a given input • Major method for prediction: regression – Many variants of regression analysis in statistics literature; not covered in this class • Neural network and k-NN a do eg essio of-the- o • SVMs for regression exist • What about trees? out- 221 Regression Trees and Model Trees • Regression tree: proposed in CART system (Breiman et al. 1984) 222 Classification and Prediction Overview • • • • • • • • • • – CART: Classification And Regression Trees – Each leaf stores a continuous-valued prediction • Average output value for the training records in the leaf • Model tree: proposed by Quinlan (1992) – Each leaf holds a regression model—a multivariate linear equation • Training: like for classification trees, but uses variance instead of purity measure for selecting split predicates Introduction Decision Trees Statistical Decision Theory Nearest Neighbor Bayesian Classification Artificial Neural Networks Support Vector Machines (SVMs) Prediction Accuracy and Error Measures Ensemble Methods 223 Classifier Accuracy Measures Predicted class True class Precision and Recall total buy_computer = yes buy_computer = no 6954 46 412 2588 3000 7366 2634 10000 buy_computer = yes buy_computer = no total 7000 • Accuracy of a classifier M, acc(M): percentage of test records that are correctly classified by M – Error rate (misclassification rate) of M = 1 – acc(M) – Given m classes, CM[i,j], an entry in a confusion matrix, indicates # of records in class i that are labeled by the classifier as class j C1 C2 C1 True positive False negative C2 False positive True negative 224 • Precision: measure of exactness – t-pos / (t-pos + f-pos) • Recall: measure of completeness – t-pos / (t-pos + f-neg) • F-measure: combination of precision and recall – 2 * precision * recall / (precision + recall) • Note: Accuracy = (t-pos + t-neg) / (t-pos + t-neg + f-pos + f-neg) 225 226 34 Limitation of Accuracy Cost-Sensitive Measures: Cost Matrix • Consider a 2-class problem PREDICTED CLASS – Number of Class 0 examples = 9990 – Number of Class 1 examples = 10 C(i|j) • If model predicts everything to be class 0, accuracy is 9990/10000 = 99.9 % ACTUAL CLASS Class=Yes Class=No Class=Yes C(Yes|Yes) C(No|Yes) Class=No C(Yes|No) C(No|No) – Accuracy is misleading because model does not detect any class 1 example • Always predicting the majority class defines the baseline C(i| j): Cost of misclassifying class j example as class i – A good classifier should do better than baseline 227 Computing Cost of Classification Cost Matrix ACTUAL CLASS Model M1 PREDICTED CLASS + ACTUAL CLASS + - + -1 100 - 1 0 Model M2 • Continuous output: it matters how far off the prediction is from the true value • Loss function: dista e et ee a d p edi ted alue – Absolute error: | y – | – Squared error: (y – 2 150 40 - 60 250 + ACTUAL CLASS • Test error (generalization error): average loss over the test set • Mean absolute error: Mean squared error: • Relative absolute error: | y(i)  y' (i) | Relative squared error:  ( y(i)  y' (i)) 2 - + 250 45 - 5 200 1 n 2   y(i)  y' (i) n i 1 1 n  | y(i)  y' (i) | n i 1 PREDICTED CLASS - + Accuracy = 80% Cost = 3910 Prediction Error Measures PREDICTED CLASS C(i|j) 228 n n i 1 i 1  | y(i)  y |  ( y(i)  y) n n i 1 Accuracy = 90% Cost = 4255 2 i 1 • Squared-error exaggerates the presence of outliers 229 Evaluating a Classifier or Predictor • Holdout method 230 Learning Curve • Accuracy versus sample size • Effect of small sample size: – The given data set is randomly partitioned into two sets • Training set (e.g., 2/3) for model construction • Test set (e.g., 1/3) for accuracy estimation – Bias in estimate – Variance of estimate – Can repeat holdout multiple times • Accuracy = avg. of the accuracies obtained • Cross-validation (k-fold, where k = 10 is most popular) – Randomly partition data into k mutually exclusive subsets, each approximately equal size – In i-th iteration, use Di as test set and others as training set – Leave-one-out: k folds where k = # of records • Expensive, often results in high variance of performance metric 231 • Helps determine how much training data is needed – Still need to have enough test and validation data to be representative of distribution 232 35 ROC (Receiver Operating Characteristic) ROC Curve • Developed in 1950s for signal detection theory to analyze noisy signals • 1-dimensional data set containing 2 classes (positive and negative) – Any point located at x > t is classified as positive – Characterizes trade-off between positive hits and false alarms • ROC curve plots T-Pos rate (y-axis) against F-Pos rate (x-axis) • Performance of each classifier is represented as a point on the ROC curve – Changing the threshold of the algorithm, sample distribution or cost matrix changes the location of the point At threshold t: TPR=0.5, FPR=0.12 233 ROC Curve 234 Diagonal Line for Random Guessing (TPR, FPR): • (0,0): declare everything to be negative class • (1,1): declare everything to be positive class • (1,0): ideal • Classify a record as positive with fixed probability p, irrespective of attribute values • Consider test set with a positive and b negative records • True positives: p*a, hence true positive rate = (p*a)/a = p • False positives: p*b, hence false positive rate = (p*b)/b = p • For every value 0p1, we get point (p,p) on ROC curve • Diagonal line: – Random guessing 235 236 Using ROC for Model Comparison • Neither model consistently outperforms the other – M1 better for small FPR – M2 better for large FPR • Area under the ROC curve – Ideal: area = 1 – Random guess: area = 0.5 237 How to Construct an ROC curve record P(+|x) True Class 1 0.95 + 2 0.93 + 3 0.87 - 4 0.85 - 5 0.85 - 6 0.85 + 7 0.76 - 8 0.53 + 9 0.43 - 10 0.25 + • Use classifier that produces posterior probability P(+|x) for each test record x • Sort records according to P(+|x) in decreasing order • Apply threshold at each unique value of P(+|x) – Count number of TP, FP, TN, FN at each threshold – TP rate, TPR = TP/(TP+FN) – FP rate, FPR = FP/(FP+TN) 238 36 How To Construct An ROC Curve Class Threshold >= + - + - + - - - + + 0.85 0.85 0.85 0.25 0.43 0.53 0.76 0.87 0.93 0.95 1.00 TP 5 4 4 3 3 2 2 1 0 FP 5 5 4 4 3 1 0 0 0 TN 0 0 1 1 2 4 5 5 5 FN 0 1 1 2 2 3 3 4 5 TPR 1 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0 FPR 1 1 0.8 0.8 0.6 0.2 0 0 0 Test of Significance • Given two models: – Model M1: accuracy = 85%, tested on 30 instances – Model M2: accuracy = 75%, tested on 5000 instances 1.0 true positive rate • Can we say M1 is better than M2? – How much confidence can we place on accuracy of M1 and M2? – Can the difference in accuracy be explained as a result of random fluctuations in the test set? ROC Curve: 0.4 0.2 0 0.2 0.4 1.0 false positive rate 239 240 Confidence Interval for Accuracy Confidence Interval for Accuracy Area = 1 -  • Classification can be regarded as a Bernoulli trial – A Be oulli t ial has possi le out o es, o e t o o g fo lassifi atio – Collection of Bernoulli trials has a Binomial distribution • Bi o ial dist i utio fo X= u e of o e tl lassified test e o ds out of • Accuracy = X / n • For large test sets (n>30), Binomial distribution is closely approximated by normal distribution with same mean and variance • Probability of getting c correct predictions if model accuracy is p (=probability to get a single prediction right): n c n c  c  p (1  p)   – E(X)=pn, Var(X)=p(1-p)n – E(ACC) = p, Var(ACC) = p(1-p) / n – ACC has a normal distribution with mean=p, variance=p(1-p)/n  P Z / 2    • Given c, or equivalently, ACC = c / n and n (#test records), can we predict p, the true accuracy of the model? • Z/2  ACC p  Z1 / 2   1    p(1  p) / n  Z1-  /2 Confidence Interval for p: 2n  ACC Z 2  Z 2  4n  ACC 4n  ACC2  /2  /2 p 2(n  Z2 / 2 ) 241 242 Comparing Performance of Two Models Confidence Interval for Accuracy • Consider a model that produces an accuracy of 80% when evaluated on 100 test instances – n = 100, ACC = 0.8 – Let 1- = 0.95 (95% confidence) – From probability table, Z/2 = 1.96 1- • Given two models M1 and M2, which is better? Z N 50 100 500 1000 5000 0.99 2.58 p(lower) 0.670 0.711 0.763 0.774 0.789 0.98 2.33 p(upper) 0.888 0.866 0.833 0.824 0.811 0.95 1.96 p 2n  ACC Z2 / 2  Z2 / 2  4n  ACC 4n  ACC2 2(n  Z2 / 2 ) 0.90 1.65 243 – – – – M1 is tested on D1 (size=n1), found error rate = e1 M2 is tested on D2 (size=n2), found error rate = e2 Assume D1 and D2 are independent If n1 and n2 are sufficiently large, then – Estimate: err1 ~ N 1 ,  1  err2 ~ N  2 ,  2  ˆ i  ei and ˆ i2  ei (1  ei ) ni 244 37 Testing Significance of Accuracy Difference An Illustrative Example • Consider random variable d = err1– err2 • Given: M1: n1 = 30, e1 = 0.15 M2: n2 = 5000, e2 = 0.25 • E[d] = |e1 – e2| = 0.1 • 2-sided test: dt = 0 versus dt  0 – Since err1, err2 are normally distributed, so is their difference – Hence d ~ N (dt, t) where dt is the true difference ˆ t2  • Estimator for dt: 0.15(1  0.15) 0.25(1  0.25)   0.0043 30 5000 – E[d] = E[err1-err2] = E[err1] – E[err2]  e1 - e2 – Since D1 and D2 are independent, variance adds up: • At 95% confidence level, Z/2 = 1.96 – At (1-) confidence level, dt  E[ d ]  Z / 2ˆ t • Interval contains zero, hence difference may not be statistically significant • But: may reject null hypothesis (dt  0) at lower confidence level dt  0.100  1.96 0.0043  0.100  0.128 e (1  e1 ) e2 (1  e2 ) ˆ t2  ˆ12  ˆ 22  1  n1 n2 245 Significance Test for K-Fold CrossValidation Classification and Prediction Overview • Each learning algorithm produces k models: – L p odu es M – L p odu es M ,M ,M , …, M k , …, M k • Both models are tested on the same test sets D1, D2,…, Dk – For each test set, compute dj = e1,j – e2,j – For large enough k, dj is normally distributed with mean dt and variance t k – Estimate: (d  d ) 2 ˆ t2   j 1 j k(k  1) d t  d  t1 ,k 1ˆ t 246 t-distribution: get t coefficient t1-,k-1 from table by looking up confidence level (1-) and degrees of freedom (k-1) • • • • • • • • • • Introduction Decision Trees Statistical Decision Theory Nearest Neighbor Bayesian Classification Artificial Neural Networks Support Vector Machines (SVMs) Prediction Accuracy and Error Measures Ensemble Methods 247 248 Ensemble Methods General Idea • Construct a set of classifiers from the training data D Step 1: Create Multiple Data Sets • Predict class label of previously unseen records by aggregating predictions made by multiple classifiers Step 2: Build Multiple Classifiers Step 3: Combine Classifiers 249 D1 D2 C1 C2 .... Original Training data Dt-1 Dt Ct -1 Ct C* 250 38 Why Does It Work? Base Classifier vs. Ensemble Error • Consider 2-class problem • Suppose there are 25 base classifiers – Each classifier has error rate  = 0.35 – Assume the classifiers are independent • Return majority vote of the 25 classifiers – Probability that the ensemble classifier makes a wrong prediction: 25 25   i  (1   ) i 13     i 25i  0.06 251 Model Averaging and Bias-Variance Tradeoff Bagging: Bootstrap Aggregation • Single model: lowering bias will usually increase variance – “ oothe odel has lo e model function well enough a ia e ut 252 • Given training set with n records, sample n records randomly with replacement ight ot Original Data Bagging (Round 1) Bagging (Round 2) Bagging (Round 3) • Ensembles can overcome this problem 1 7 1 1 2 8 4 8 3 10 9 5 4 8 1 10 5 2 2 5 6 5 3 5 7 10 2 9 8 10 7 6 9 5 3 3 10 9 2 7 • Train classifier for each bootstrap sample • Note: each training record has probability 1 – (1 – 1/n)n of being selected at least once in a sample of size n 1. Let models overfit • Low bias, high variance 2. Take care of the variance problem by averaging many of these models • This is the basic idea behind bagging 253 Bagged Trees 254 Typical Result • Create k trees from training data – Bootstrap sample, grow large trees • Design goal: independent models, high variability between models • Ensemble prediction = average of individual tree predictions (or majority vote) • Works the same way for other classifiers (1/k)· + (1/k)· +…+ (1/k)· 255 256 39 Typical Result Typical Result 257 258 Bagging Challenges Additive Grove • Ideal case: all models independent of each other • Train on independent data samples • Ensemble technique for predicting continuous output • Instead of individual trees, train additive models – Prediction of single Grove model = sum of tree predictions – Problem: limited amount of training data • Prediction of ensemble = average of individual Grove predictions • Combines large trees and additive models • Training set needs to be representative of data distribution – Bootst ap sa pli g allo s eatio of independent training sets a al ost – Challenge: how to train the additive models without having the first trees fit the training data too well • Diversify models, because similar sample might result in similar tree • Next tree is trained on residuals of previously trained trees in same Grove model • If previously trained trees capture training data too well, next tree is mostly trained on noise – Random Forest: limit choice of split attributes to small random subset of attributes (new selection of subset for each node) when training tree – Use different model types in same ensemble: tree, ANN, SVM, regression models +…+ (1/k)· + (1/k)· +…+ +…+ (1/k)· +…+ 259 260 Training Groves Typical Grove Performance 10 • Root mean squared error 9 – Lower is better 8 + + + + + • Horizontal axis: tree size + 7 – Fraction of training data when to stop splitting 6 + + + • Vertical axis: number of trees in each single Grove model • 100 bagging iterations 5 4 0.13 3 2 1 0.5 261 0.2 0.1 0.05 0.02 0.01 0.005 0.002 0 262 40 Boosting Boosting • Iterative procedure to adaptively change distribution of training data by focusing more on previously misclassified records – Initially, all n records are assigned equal weights – Record weights may change at the end of each boosting round • Records that are wrongly classified will have their weights increased • Records that are classified correctly will have their weights decreased Original Data Boosting (Round 1) Boosting (Round 2) Boosting (Round 3) 1 7 5 4 2 3 4 4 3 2 9 8 4 8 4 10 5 7 2 4 6 9 5 5 7 4 1 4 8 10 7 6 9 6 4 3 10 3 2 4 • Assume record 4 is hard to classify • Its weight is increased, therefore it is more likely to be chosen again in subsequent rounds 263 264 Example: AdaBoost AdaBoost Details   w(ji )  i if Ci ( x j )  y j  1   i Zi  1 if Ci ( x j )  y j  where Zi is the normalizat ion factor • Weight update: • Base classifiers: C1, C2,…, CT • Error rate (n training records, wj are weights that sum to 1):  i   wj Ci ( x j )  y j  wj ( i 1)  • Weights initialized to 1/n • Zi ensures that weights add to 1 • If any intermediate rounds produce error rate higher than 50%, the weights are reverted back to 1/n and the resampling procedure is repeated • Final classification: T n j 1 • Importance of a classifier: 1 i    i  ln   i  C * ( x)  arg max   i Ci ( x)  y i 1 y 265 266 Illustrating AdaBoost Illustrating AdaBoost B1 Initial weights for each data point Original Data 0.1 0.1 0.1 +++ - - - - - ++ 0.0094 Data points for training Boosting Round 1 +++ 0.0094 0.4623 - - - - - - -  = 1.9459 B2 Boosting Round 2 0.3037 - - - 0.0009 0.0422 - - - - - ++ 0.1819 0.0038  = 2.9323 New weights B1 0.0094 Boosting Round 1 +++ B3 0.0094 0.4623 - - - - - - -  = 1.9459 Note: The numbers appear to be wrong, but they convey the right idea… 267 0.0276 Boosting Round 3 +++ ++ ++ + ++ Overall +++ - - - - -  = 3.8744 ++ Note: The numbers appear to be wrong, but they convey the right idea… 268 41 Bagging vs. Boosting • Analogy • Sampling procedure • Overfitting Classification/Prediction Summary – Baggi g: diag osis ased o ultiple do to s ajo it ote – Boosti g: eighted ote, ased o do to s p e ious diag osis a u a – Bagging: records have same weight; easy to train in parallel – Boosting: weights record higher if model predicts it wrong; inherently sequential process – Bagging robust against overfitting – Boosting susceptible to overfitting: make sure individual models do not overfit • Accuracy usually significantly better than a single classifier • Additive Grove – Best boosted model often better than best bagged model – Combines strengths of bagging and boosting (additive models) – Shown empirically to make better predictions on many data sets – Training more tricky, especially when data is very noisy 269 • Forms of data analysis that can be used to train models from data and then make predictions for new records • Effective and scalable methods have been developed for decision tree induction, Naive Bayesian classification, Bayesian networks, rule-based classifiers, Backpropagation, Support Vector Machines (SVM), nearest neighbor classifiers, and many other classification methods • Regression models are popular for prediction. Regression trees, model trees, and ANNs are also used for prediction. 270 Classification/Prediction Summary • K-fold cross-validation is a popular method for accuracy estimation, but determining accuracy on large test set is equally accepted – If test sets are large enough, a significance test for finding the best model is not necessary • Area under ROC curve and many other common performance measures exist • Ensemble methods like bagging and boosting can be used to increase overall accuracy by learning and combining a series of individual models – Often state-of-the-art in prediction quality, but expensive to train, store, use • No single method is superior over all others for all data sets – Issues such as accuracy, training and prediction time, robustness, interpretability, and scalability must be considered and can involve trade-offs 271 42