Search engines have focused on solving the document retrieval problem, so their scoring functions do not handle naturally non-traditional IR data types, such as numerical or categorical. Therefore, on domains beyond traditional search, scores representing strengths of associations or matches may vary widely. As such, the original model doesn’t suffice, so relevance ranking is performed as a two-phase approach with 1) regular search 2) external model to re-rank the filtered items. Metrics such as click-through and conversion rates are associated with the users’ response to items served. The predicted selection rates that arise in real-time can be critical for optimal matching. For example, in recommender systems, predicted performance of a recommended item in a given context, also called response prediction, is often used in determining a set of recommendations to serve in relation to a given serving opportunity. Similar techniques are used in the advertising domain. To address this issue the authors have created ML-Scoring, an open source framework that tightly integrates machine learning models into a popular search engine (SOLR/Elasticsearch), replacing the default IR-based ranking function. A custom model is trained through either Weka or Spark and it is loaded as a plugin used at query time to compute custom scores.
2. Presenters
Diana Hu
Senior Data Scientist
@sdianahu
diana.hu@verizon.com
Joaquin Delgado, PhD.
Director of Engineering
@joaquind
joaquin.a.delgado@verizon.com
3. Disclaimer
The content of this presentation are of the
authors’ personal statements and does not
officially represent their employer’s view in
anyway. Included content is especially not
intended to convey the views of OnCue or Verizon.
4. Index
1. Introduction
1. What to expect?
2. Scaling recommender systems is hard
2. Recommender System Problem as a Search Problem
1. Representing queries as recommendations
3. Introduction to Search and Information Retrieval
1. Scalability in search
2. Introduction to Elasticsearch
4. Overview of Machine Learning Techniques for Recommender Systems
1. Learning to rank
2. Scalability in machine learning
3. ML software frameworks
5. Re-writing the ranking function
1. Writing a new ranking/scoring function in Elasticsearch
2. Training a spark model as a Elasticsearch plugin for custom ranking/scoring function
6. References
6. What to expect from this tutorial?
• The focus is on practical examples of how
to implement scalable recommender
systems using search and learning-to-rank
(machine learning) techniques
• What it is not
• Deep dive into any specific areas (Search,
RecSys, Learning to rank, or Machine learning)
• Algorithmic survey
• Comparative Analysis
8. What is a recommendation?
Beyond rating prediction
9. Paradigms of recommender systems
• Reduce information load by estimating
relevance
• Ranking Approaches:
• Collaborative filtering: “Tell me what is popular
amongst my peers”
• Content Based: “Show me more of what I liked”
• Knowledge Based: “Tell me what fits my needs”
• Hybrid
10. Model Type Pros Cons
Collaborative • No metadata engineering
effort
• Serendipity of results
• Learns market segments
• Requires rating feedback
• Cold start for new users and
new items
Content-based • No community required
• Comparison between
items possible
• Content descriptions
necessary
• Cold start for new users
• No serendipity
Knowledge-
based
• Deterministic
• Assured quality
• No cold-start
• Interactive user sessions
• Knowledge engineering
effort to bootstrap
• Static
• Does not react to short-term
trends
11. Scaling recommender systems is hard!
• Millions of users
• Millions of items
• Cold start for ever increasing size of
catalog and new users added
• Imbalanced Datasets – power law
distribution is quite common
• Many algorithms have not been fully tested
at “Internet Scale”
13. Content-based methods inspired by IR
• Rec Task: Given a user profile find the best matching
items by their attributes
• Similarity calculation: based on keyword overlap
between user/items
• Neighborhood method (i.e. nearest neighbor)
• Query-based retrieval (i.e Rocchio’s method)
• Probabilistic methods (classical text classification)
• Explicit decision models
• Feature representation: based on content analysis
• Vector space model
• TF-IDF
• Topic Modeling
14. Search queries as content-based
recommendations
• Exact matching (Boolean)
• Relevant or not relevant (no ranking)
• Ranking by similarity to query (Vector
Space Model)
• Text similarity: Bag of words, TF-IDF, Incidence
Matrix
• Ranking by importance (e.g. PageRank)
15. Content-based similarity measures
• Simple match
• Dice’s Coefficient
• Jaccard’s Coefficient
• Cosine Coefficient
• Overlap Coefficient
3D Term Vector Space
16. Knowledge-based methods inspired by IR
• Rec Task: Given explicit recommendation rules find the
best matches between user’s requirements and item’s
characteristics (i.e., which item should be recommended in
which context?)
• Similarity calculation: based on constraint satisfaction
problem and distance similarity requirements<->attributes
• Conjunctive queries
• Similarity metrics for item retrieval
• Feature representation: based on query representation
• User defined preferences
• Utility-based preferences
• Conjoint analysis
17. Search queries as knowledge-based
recommendations
• Constraint satisfaction problem (CSP) is a tuple
(V,D,C)
• V – set of variables
• D – set of finite domains for V
• C – set of constraints of possible V permutations
• Recommendation as CSP:
(V,D,C) => (Vi U Vu, D, Cr U Ci U Cf U REQ)
• Vu – user properties (possible user’s requirements)
• Vi – item properties
• Cr – compatibility constraints (possible Vc permutations)
• Ci – Item constraints (conjunction fully defines an item)
• Cf – filter conditions (define Vu<->Vi relationships)
• REQ – user’s requirements
19. Search
Search is about finding specific things that are either
known or assumed to exist, Discovery is about is
about helping the user encounter what he/she didn’t
even know exists
Both Search and Discovery can be achieved through
a query based data/information system.
Predicate Logic and Declarative Languages Rock!
20. Examples of query based systems
• Focused on Search
• Search engines
• Database systems
• Focus on Discovery
• Recommender systems
• Advertising systems
21. IR: The science behind search!
Information Retrieval (IR) is a query based on
data retrieval + relevance ranking (scoring)
usually applied to unstructured data (i.e. text
documents and fields); often referred to as full-
text or keyword search.
Have you heard of Bag-of-Words?
Vector Space Representation?
What about TF-IDF?
23. Retrieval Models
Model Type Query
Representation
Document
Representation
Retrieval
Boolean • Boolean
expressions
• Connected by
AND, OR, NOT
• Set of keywords
• Bag of words
• Binary term weight
• Exact match
• Binary relevance
• No ranking
Vector
Space
Model
• Vector
• Desired terms
with optional
weights
• Vectors
• Bag of words with
weight based on
TF-IDF scheme
• Similarity score
• Output documents
are ranked
• Relevance
feedback support
Probabilisti
c
• Similarity with
priors
• Document
relevance
• Ranks documents
in decreasing
probability of
relevance
25. Search Engines: the big hammer!
• Search engines are largely used to solve
non-IR search problems, and here is why:
• Widely available
• Fast and scalable distributed systems
• Integrates well with existing data stores (SQL and NoSQL)
26. But are we using the right tool?
• Search Engines were originally designed
for IR.
• More complex non-IR search/discovery
tasks sometimes require a multi-phase,
multi-system approach
28. Elasticsearch
• What is Elasticsearch?
• Elasticsearch is an open-source search engine
• Elasticsearch is written in Java
• Built on top of Apache Lucene™
• A distributed real-time document store where every field is
indexed and searchable out-of-the box
• A distributed search engine with real-time analytics
• Has a plugin architecture that facilitates extending the
core system
• Written with NRT and cloud support in mind
• Easy index, shard and replicas creation on live cluster
• Has Optimistic Concurrency Control
29. Examples of scaling challenges
• More than 50 millions of documents a day
• Real time search
• Less than 200ms average query latency
• Throughput of at least 1000 QPS
• Multilingual indexing
• Multilingual querying
30. Who uses ES?
• Wikipedia
• Uses Elasticsearch to provide full-text search with highlighted
search snippets, and search-as-you-type and did-you-mean
suggestions.
• The Guardian
• Uses Elasticsearch to combine visitor logs with social -network
data to provide real-time feedback to its editors about the
public’s response to new articles.
• Stack Overflow
• Combines full-text search with geo-location queries and uses
more-like-this to find related questions and answers.
• GitHub
• Uses Elasticsearch to query 130 billion lines of code.
31. How ES scales?
• Sharding and Replicas
• Several indices (at least one index for each day of
data)
• Indices divided into multiple shards
• Multiple replicas of a single shard
• Real-time, synchronous replication
• Near-real-time index refresh (1 to 30 seconds)
33. Querying ES
Node 1 Node 2 Node 3 Node 4
Node 5 Node 6 Node 7 Node 8
ES Index
Application
34. Using Search Engines for RS
• Its not just about rating prediction and ranking
• Business filtering logic
• Age restrictions
• Catalog navigation context (e.g. e-commerce)
• Promotional materials
• Low latency and scale
• SLAs on response times including query, responses
and presentation
• Actual time for computing recommendations is just a
small fraction of total allocated time
35. Stacking things up
Visualization / UI
Retrieval
Ranking
Query Generation and
Contextual Pre-filtering
Model Building
Index Building
Data/Events Collections
Data Analytics
Contextual Post Filtering
OnlineOffline
Experimentation
37. 4. Overview of Machine Learning
Techniques for Recommender Systems
38. Machine Learning
Machine Learning in particular supervised learning
refer to techniques used to learn how to classify or
score previously unseen objects based on a training
dataset
Inference and Generalization are the Key!
39. Recommendations as data mining
Amatriain, Xavier, et al.
"Data mining methods for
recommender systems."
Recommender Systems
Handbook. Springer US,
2011. 39-71.
40. Learning to rank
• Formulate the problem as standard
supervised learning
• Training data can be cardinal or binary
• Various approaches:
• Pointwise: Typically approximated by regression
• Pairwise: Approximated via binary classifier
• Listwise: Directly optimize whole list (difficult!)
• A trick with ES is to include raw scores
returned by ES into the feature vector
41. Learning to rank with ES
Elastic Search
ES
Query
ES
Index
Input:
Contextual features
Potential
Matches
Trained
Ranking
Model
ML
Framework
+
Gold
Dataset
Output:
Ranked Results
42. Web scale ML challenges
• Massive amount of examples
• Billions of features
• Big models don’t fit in a single machine’s memory
• Variety of algorithms that need to be scaled up
A Note of Caution….
43. “Invariably, simple models and
a lot of data trump more elaborate
models based on less data.”
Alon Halevy, Peter Norvig, and
Fernando Pereira, Google
http://static.googleusercontent.com/external_content/untrusted_dlcp/research.google.com/en//pubs/archive/35179.pdf
44. Scalability in Machine Learning
• Distributed systems – Fault tolerance,
Throughput vs. latency
• Parallelization Strategies – Hashing, trees
• Processing – Map reduce variants, MPI,
graph parallel
• Databases – Key/Value Stores, NoSQL
45. What is Spark?
Fast, expressive cluster computing system
45
BlinkDB
approx queries
Spark SQL
structured data
MLlib
machine
learning
Spark
Streaming
real-time
GraphX
graph
Analytics
Spark Core
46. What is Spark?
• Work on distributed collections like local ones
• RDD:
• Immutable
• Parallel transforms
• Resilient and configurable persistence
• Operations
• Transforms: Lazy operations (map, filter, join,…)
• Actions: Return/write results (collect, save, count,…)
47. ML Software Framework: Spark MLlib
• Subproject with ML primitives
• Building blocks (as a framework vs. library)
• Large scale statistics
• Classification
• Regression
• Clustering
• Matrix factorization
• Optimization
• Frequent pattern mining
• Dimensionality reduction
48. What is ML-Scoring?
• Creates an Elastic Search (ES) document index of
instances
• Trains a supervised learning ML model from a dataset
of instances + labels
• Generate an Elasticsearch plugin that uses the trained
ML model to score documents at query time
• A
•
An Open Source POC!
49. Remember the elephant?
Visualization / UI
Retrieval
Ranking
Query Generation and
Contextual Pre-filtering
Model Building
Index Building
Data/Events Collections
Data Analytics
Contextual Post Filtering
OnlineOffline
Experimentation
50. Simplifying the Stack!
Visualization / UI
Query Generation and
Contextual Pre-filtering
Model Building
Index Building
Data/Events Collections
Data Analytics
Retrieval
Contextual Post Filtering
Ranking
OnlineOffline
Experimentation
53. Using ML-Scoring
• Creating an ES Index
• Boolean queries
• More-Like-This queries
• Built-in scoring functions
• Scoring script
• Scoring plugin
• ML-Score evaluator using Spark
• ML-Score query
54. Creating an Index in ES
POST /my_movie_catalog/movies/_bulk
{ "index": { "_id": 1 }}
{ ”genre" : “Documentary”, ”productID" : "XHDK-A-1293-#fJ3" , “title” :
“Olympic Sports”, “content” : “Olympic greateness…“, price” : 20}
{ "index": { "_id": 2 }}
{ ”genre" : “Sports”, ”productID" : "KDKE-B-9947-#kL5", “title” : “NY
Yankees: Winning the World Series”, , “content” : “There is no better
team than the NY…“ “price” :20}
{ "index": { "_id": 3 }}
{ ”genre" : “Action”, “productID" : "JODL-X-1937-#pV7",”title” :
“Rambo III”, , “content” : “Sylvester Stallone is evermore…“ “price” :
18}
{ "index": { "_id": 4 }}
{ ”genre" : “Children”, ”productID" : "QQPX-R-3956-#aD8", “title” :
“Fairy Tale”, , “content” : “Once upon a time…“, “price” : 30}
55. Boolean queries
• SQL representation
SELECT movie
FROM movies
WHERE (price = 20 OR productID = "XHDK-A-1293-#fJ3")
AND (price != 30)
• ES DSL
GET /my_movie_catalog/movies/_search
{
"query" : {
"filtered" : {
"filter" : {
"bool" : {
"should" : [
{ "term" : {"price" : 20}},
{ "term" : {"productID" : "XHDK-A-1293-#fJ3"}}
],
"must_not" : {
"term" : {"price" : 30}
…
56. Content based similarity queries (MLT)
{
"more_like_this" : {
"fields" : ["title", "description"],
"like_text" : "Once upon a time",
"min_term_freq" : 1,
"max_query_terms" : 12
}
}
• The More Like This Query (MLT Query) finds documents
that are "like" a given set of documents. In order to do so,
MLT selects a set of representative terms of these input
documents, forms a query using these terms, executes the
query and returns the results.
64. Limitations of ranking using ES
practical scoring function
• Stateless computation
• Meant primarily for text search
• Hard to represent context and history
• Limited complexity (simple math functions only)
• Nevertheless, original score should not be
discarded as it may become handy!
65. Scoring plugin in ES
public class PredictorPlugin extends AbstractPlugin {
@Override
public String name() {
return getClass().getName();
}
@Override
public String description() {
return "Simple plugin to predict values.";
}
public void onModule(ScriptModule module) {
module.registerScript(
PredictorScoreScript.SCRIPT_NAME,
PredictorScoreScript.Factory.class);
}
}
66. ML-Scoring evaluator using Spark
class SparkPredictorEngine[M](val readPath: String, val spHelp:
SparkModelHelpers[M]) extends PredictorEngine {
private var _model: ModelData[M] = ModelData[M]()
override def getPrediction(values: Collection[IndexValue]) = {
if (_model.clf.nonEmpty) {
val v = ReadUtil.cIndVal2Vector( values, _model.mapper)
_model.clf.get.predict(v)
} else {
throw new PredictionException("Empty model");
}
}
def readModel() = _model = spHelp.readSparkModel(readPath)
def getModel: ModelData[M] = _model
69. Potential issues
• Performance
• It may be a problem if the search space is very
large and/or the computation to intensive
• Operations
• Code running on a key infrastructure
• Versioning and binary compatibility
70. Summary
• Importance of the whole picture – RS seen from the lenses
of the whole elephant
• RS research is a new field in comparison to IR
• Scalability is hard! Why not learn from all of RS’s cousins:
• Search
• Distributed systems
• Databases
• Machine learning
• Content analysis
• …
• Bridging the gap between research and engineering is an
ongoing effort
71. References
• Baeza-Yates, R., & Ribeiro-Neto, B. 2011. Modern information retrieval. New York:
ACM press.
• Chirita, P. A., Firan, C. S., & Nejdl, W. 2007. Personalized query expansion for the web.
In Proceedings of the 30th annual international ACM SIGIR conference on Research
and development in information retrieval (pp. 7-14). ACM.
• Croft, W. B., Metzler, D., & Strohman, T. 2010. Search engines: Information retrieval in
practice. Reading: Addison-Wesley.
• Dunning, T. 1993. Accurate methods for the statistics of surprise and
coincidence. Computational linguistics, 19(1), 61-74.
• Elastic, Elasticsearch: RESTful, Distributed Search & Analytics. 2015.
https://www.elastic.co/products/elasticsearch.
• Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., & Witten, I. H. 2009.
The WEKA data mining software: an update. ACM SIGKDD explorations
newsletter, 11(1), 10-18.
• Ihaka, R., & Gentleman, R. 1996. R: a language for data analysis and graphics. Journal
of computational and graphical statistics, 5(3), 299-314.
72. References
• Kantor, P. B., Rokach, L., Ricci, F., & Shapira, B. 2011. Recommender systems handbook.
Springer.
• Manning, C. D., Raghavan, P., & Schütze, H. 2008. Introduction to information retrieval.
Cambridge: Cambridge university press.
• Qiu, F., & Cho, J. 2006. Automatic identification of user interest for personalized search.
In Proceedings of the 15th international conference on World Wide Web (pp. 727-736).
ACM.
• Sun, J. T., Zeng, H. J., Liu, H., Lu, Y., & Chen, Z. 2005. Cubesvd: a novel approach to
personalized web search. In Proceedings of the 14th international conference on World
Wide Web (pp. 382-390). ACM.
• Xing, B., & Lin, Z. 2006. The impact of search engine optimization on online advertising
market. In Proceedings of the 8th international conference on Electronic commerce: The
new e-commerce: innovations for conquering current barriers, obstacles and limitations to
conducting successful business on the internet (pp. 519-529). ACM.
• Zaharia, M., Chowdhury, M., Franklin, M. J., Shenker, S., & Stoica, I. 2010. Spark: cluster
computing with working sets. In Proceedings of the 2nd USENIX conference on Hot topics in
cloud computing (Vol. 10, p. 10)
73. Additional Credits
• Doug Kang
• Data Scientist, Verizon OnCue
• Federico Ponte
• System Engineer from Mahisoft
• Yessika Labrador
• Data Engineer from Mahisoft
Editor's Notes
Real-life recommender systems: How it works in the industry, outside of academics settings
Rating prediction
Ranking
Similarity
Filtering
UI/Presentation
cannot escape our own perspectives
Which is why, if we want to do good work—and particularly if we want to innovate—we need to have ‘other.’
why the only way to see our biases is through other people.
-trunk - snake
-tusk - sword
-stomach - whale
-tail - reed
-leg - tree
the elephant in the room is truth
KN: items such as apartments and cars are not purchased very often, therefore rating-based systems often do not perform well/ For example, only financial services must be recommended that support the investment period
Diana
Similarity calculation aka similarity-based retrieval
Rocchio’s relevance feedback in 1960
Challenge to deal with unsatiesdiable REQ and empty result sets
User def- explicit
Utility scoring rules predifined
Conjoint- analysis of of past interaction
Geo sim
Database Systems
Data retrieval with no scoring
AI Systems
Games: e.g. Chess – possible next moves with associated move quality score
E-Commerce Systems
Sorting can be based of relevance score, proximity (geo-location), price, etc.
Recommender Systems
Taste scoring based on content and/or user similarity
Advertising Systems
Two-way search (matching) and optimization problem with scores modeled as bids
Pro Boolean:
Popular retrieval model because:
Easy to understand for simple queries.
Clean formalism.
Boolean models can be extended to include ranking.
Reasonably efficient implementations possible for normal queries
Cons Booelan:
Very rigid: AND means all; OR means any.
Difficult to express complex user requests.
Difficult to control the number of documents retrieved.
All matched documents will be returned.
Difficult to rank output.
All matched documents logically satisfy the query.
Difficult to perform relevance feedback.
If a document is identified by the user as relevant or irrelevant, how should the query be modified?
Pro VSP
Similarity based on occurrence frequencies of keywords in query and document.
Automatic relevance feedback can be supported:
Relevant documents “added” to query.
Irrelevant documents “subtracted” from query.
Cons VSP:
How to determine important words in a document?
Word sense?
Word n-grams (and phrases, idioms,…) terms (old school)
How to determine the degree of importance of a term within a document and within the entire collection?
How to determine the degree of similarity between a document and the query?
In the case of the web, what is the collection and what are the effects of links, formatting information, etc.?
The inner product is unbounded.
Favors long documents with a large number of unique terms.
Measures how many terms matched but not how many terms are not matched.
Explain what an index is
Search engines are widely used to solve non-IR search problems and here is why:
Available: popular open source search engines (e.g. Apache SOLR and Elastic Search) based on a mature search library (Apache Lucene)
Fast and Scalable : distributed (inverted) indexing and retrieval via scattering/gathering techniques
Integrates well with existing data stores (SQL and No-SQL)
They implement text-based scoring, including fuzzy match and some variation of TF-IDF or Okapi BM25, etc.
Filter using querying a search engine
Rank results based on a pre-generated ML predictive model
Search engines are not ACID databases. By nature they are not transactional become eventually consistent.
There is more to Recsys than algorithms and ranking
- Retrieval
- User Interface & Feedback
- Data
- AB Testing
- Systems & Architectures
Machine Learning can allow learning a user model or profile of a particular user based on:
Sample interaction
Rated examples
This model or profile can then be used to:
Recommend items
Filter information
Predict behavior
Popularity is the obvious baseline
Ratings prediction is a clear secondary data input
that allows for personalization
1. Pointwise
- Ranking function minimizes loss function defined on
individual relevance judgment
- Ranking score based on regression or classification
- Ordinal regression, Logistic regression, SVM, GBDT, …
2. Pairwise
-Loss function is defined on pair-wise preferences
-Goal: minimize number of inversions in ranking
-Ranking problem is then transformed into the binary
classification problem
-RankSVM, RankBoost, RankNet, Frank…
3.Listwise
- Indirect Loss Function
− RankCosine: similarity between ranking list
and ground truth as loss function
− ListNet: KL-divergence as loss function by
defining a probability distribution
− Problem: optimization of listwise loss function
may not optimize IR metrics
- Directly optimizing IR measures (difficult since they
are not differentiable)
You write a single program similar to DryadLINQ
Distributed data sets with parallel operations on them are pretty standard; the new thing is that they can be reused across ops
Variables in the driver program can be used in parallel ops; accumulators useful for sending information back, cached vars are an optimization
Mention cached vars useful for some workloads that won’t be shown here
Mention it’s all designed to be easy to distribute in a fault-tolerant fashion
Basic statistics
summary statistics
correlations
stratified sampling
hypothesis testing
random data generation
Classification and regression
linear models (SVMs, logistic regression, linear regression)
naive Bayes
decision trees
ensembles of trees (Random Forests and Gradient-Boosted Trees)
isotonic regression
Collaborative filtering
alternating least squares (ALS)
Clustering
k-means
Gaussian mixture
power iteration clustering (PIC)
latent Dirichlet allocation (LDA)
streaming k-means
Dimensionality reduction
singular value decomposition (SVD)
principal component analysis (PCA)
Feature extraction and transformation
Frequent pattern mining
FP-growth
association rules
PrefixSpan
Evaluation metrics
PMML model export
Optimization (developer)
stochastic gradient descent
limited-memory BFGS (L-BFGS)
There is more to Recsys than algorithms and ranking
- Retrieval
- User Interface & Feedback
- Data
- AB Testing
- Systems & Architectures
There is more to Recsys than algorithms and ranking
- Retrieval
- User Interface & Feedback
- Data
- AB Testing
- Systems & Architectures
Performance
It may be a problem if the search space is very large and/or the computation to intensive
Operations
Code running on a key infrastructure
People are more hesitant to touch an infrastructure/DB component such as Elasticsearch. Similar concerns exist surrounding DB stored procedures.
No way to sandbox a native plugin
Requires strong automated regression and performance testing
How handle versioning and binary compatibility
Potential deployment issues
Upgrades to Elastic search, the plugin code and/or the models may present challenges