The library is a part of the ecosystem:

• ml_algo library - implementation of popular machine learning algorithms
• ml_preprocessing library - a library for data preprocessing
• ml_linalg library - a library for linear algebra
• ml_dataframe library- a library for storing and manipulating data

Table of contents

• What is ml_algo for
• The library content
• Examples
  • Logistic regression
  • Linear regression
  • Decision tree-based classification
  • KDTree-based data retrieval
• Models retraining
• Notes on gradient-based optimisation algorithms
• Helpful articles on algorithms standing behind the library
• Contacts

The main purpose of the library is to give native Dart implementation of machine learning algorithms to those who are interested both in Dart language and data science. This library aims at Dart VM and Flutter. It is also possible to use its core features in web applications using web assembly.

• Model selection

  • CrossValidator. A factory that creates instances of cross validators. Cross-validation allows researchers to fit different hyperparameters of machine learning algorithms assessing prediction quality on different parts of a dataset.

• Classification algorithms

  • LogisticRegressor. A class that performs linear binary classification of data. To use this kind of classifier your data has to be linearly separable.

    • LogisticRegressor.SGD. Implementation of the logistic regression algorithm based on stochastic gradient descent with L2 regularisation. To use this kind of classifier your data has to be linearly separable.

    • LogisticRegressor.BGD. Implementation of the logistic regression algorithm based on batch gradient descent with L2 regularisation. To use this kind of classifier your data has to be linearly separable.

    • LogisticRegressor.newton. Implementation of the logistic regression algorithm based on Newton-Raphson method with L2 regularisation. To use this kind of classifier your data has to be linearly separable.

  • SoftmaxRegressor. A class that performs linear multiclass classification of data. To use this kind of classifier your data has to be linearly separable.

  • DecisionTreeClassifier A class that performs classification using decision trees. May work with data with non-linear patterns.

  • KnnClassifier A class that performs classification using k nearest neighbours algorithm - it makes predictions based on the first k closest observations to the given one.

• Regression algorithms

  • LinearRegressor. A general class for finding a linear pattern in training data and predicting outcomes as real numbers.

    • LinearRegressor.lasso Implementation of the linear regression algorithm based on coordinate descent with lasso regularisation

    • LinearRegressor.SGD Implementation of the linear regression algorithm based on stochastic gradient descent with L2 regularisation

    • LinearRegressor.BGD Implementation of the linear regression algorithm based on batch gradient descent with L2 regularisation

    • LinearRegressor.newton Implementation of the linear regression algorithm based on Newton-Raphson method with L2 regularisation

  • KnnRegressor A class that makes predictions for each new observation based on the first k closest observations from training data. It may catch non-linear patterns of the data.

• Clustering and retrieval algorithms

  • KDTree An algorithm for efficient data retrieval.
  • Locality sensitive hashing. A family of algorithms that randomly partition all reference data points into different bins, which makes it possible to perform efficient K Nearest Neighbours search, since there is no need to search for the neighbours through the entire data. The family is represented by the following classes:
    • RandomBinaryProjectionSearcher

For more information on the library's API, please visit the API reference

Let's classify records from a well-known dataset - Pima Indians Diabetes Database via Logistic regressor

Important note:

Please pay attention to problems that classifiers and regressors exposed by the library solve. For e.g., Logistic regressor solves only binary classification problems, and that means that you can't use this classifier with a dataset with more than two classes, keep that in mind - in order to find out more about regressors and classifiers, please refer to the API documentation of the package

Import all necessary packages. First, it's needed to ensure if you have ml_preprocessing and ml_dataframe packages in your dependencies:

dependencies:
  ml_dataframe: ^1.5.0
  ml_preprocessing: ^7.0.2

We need these repos to parse raw data in order to use it further. For more details, please visit ml_preprocessing repository page.

Important note:

Regressors and classifiers exposed by the library do not handle strings, booleans and nulls, they can only deal with numbers! You necessarily need to convert all the improper values of your dataset to numbers, please refer to ml_preprocessing library to find out more about data preprocessing.

import 'package:ml_algo/ml_algo.dart';
import 'package:ml_dataframe/ml_dataframe.dart';
import 'package:ml_preprocessing/ml_preprocessing.dart';

We have 2 options here:

• Download the dataset from Pima Indians Diabetes Database.

final samples = await fromCsv('datasets/pima_indians_diabetes_database.csv');

flutter:
  assets:
    - assets/datasets/pima_indians_diabetes_database.csv

import 'package:flutter/services.dart' show rootBundle;
import 'package:ml_dataframe/ml_dataframe.dart';

void main() async {
  final rawCsvContent = await rootBundle.loadString('assets/datasets/pima_indians_diabetes_database.csv');
  final samples = DataFrame.fromRawCsv(rawCsvContent);
}

• Or we may simply use getPimaIndiansDiabetesDataFrame function from ml_dataframe package. The function returns a ready to use DataFrame instance filled with Pima Indians Diabetes Database data.

import 'package:ml_dataframe/ml_dataframe.dart';

void main() {
  final samples = getPimaIndiansDiabetesDataFrame();
}

Data in this file is represented by 768 records and 8 features. The 9th column is a label column, it contains either 0 or 1 on each row. This column is our target - we should predict a class label for each observation. The column's name is Outcome. Let's store it:

final targetColumnName = 'Outcome';

Now it's the time to prepare data splits. Since we have a smallish dataset (only 768 records), we can't afford to split the data into just train and test sets and evaluate the model on them, the best approach in our case is Cross-Validation. According to this, let's split the data in the following way using the library's splitData function:

final splits = splitData(samples, [0.7]);
final validationData = splits[0];
final testData = splits[1];

splitData accepts a DataFrame instance as the first argument and ratio list as the second one. Now we have 70% of our data as a validation set and 30% as a test set for evaluating generalization errors.

Then we may create an instance of CrossValidator class to fit the hyperparameters of our model. We should pass validation data (our validationData variable), and a number of folds into CrossValidator constructor.

final validator = CrossValidator.kFold(validationData, numberOfFolds: 5);

Let's create a factory for the classifier with desired hyperparameters. We have to decide after the cross-validation if the selected hyperparameters are good enough or not:

final createClassifier = (DataFrame samples) =>
  LogisticRegressor(
    samples
    targetColumnName,
  );

If we want to evaluate the learning process more thoroughly, we may pass collectLearningData argument to the classifier constructor:

final createClassifier = (DataFrame samples) =>
  LogisticRegressor(
    ...,
    collectLearningData: true,
  );

This argument activates collecting costs per each optimization iteration, and you can see the cost values right after the model creation.

Assume, we chose perfect hyperparameters. In order to validate this hypothesis, let's use CrossValidator instance created before:

final scores = await validator.evaluate(createClassifier, MetricType.accuracy);

Since the CrossValidator instance returns a Vector of scores as a result of our predictor evaluation, we may choose any way to reduce all the collected scores to a single number, for instance, we may use Vector's mean method:

final accuracy = scores.mean();

Let's print the score:

print('accuracy on k fold validation: ${accuracy.toStringAsFixed(2)}');

We can see something like this:

accuracy on k fold validation: 0.75

Let's assess our hyperparameters on the test set in order to evaluate the model's generalization error:

final testSplits = splitData(testData, [0.8]);
final classifier = createClassifier(testSplits[0]);
final finalScore = classifier.assess(testSplits[1], MetricType.accuracy);

The final score is like:

print(finalScore.toStringAsFixed(2)); // approx. 0.75

If we specified collectLearningData parameter, we may see costs per each iteration in order to evaluate how our cost changed from iteration to iteration during the learning process:

print(classifier.costPerIteration);

Seems, our model has a good generalization ability, and that means we may use it in the future. To do so we may store the model in a file as JSON:

await classifier.saveAsJson('diabetes_classifier.json');

After that we can simply read the model from the file and make predictions:

import 'dart:io';

void main() {
  // ...
  final fileName = 'diabetes_classifier.json';
  final file = File(fileName);
  final encodedModel = await file.readAsString();
  final classifier = LogisticRegressor.fromJson(encodedModel);
  final unlabelledData = await fromCsv('some_unlabelled_data.csv');
  final prediction = classifier.predict(unlabelledData);

  print(prediction.header); // ('class variable (0 or 1)')
  print(prediction.rows); // [ 
                        //   (1),
                        //   (0),
                        //   (0),
                        //   (1),
                        //   ...,
                        //   (1),
                        // ]
  // ...
}

Please note that all the hyperparameters that we used to generate the model are persisted as the model's read-only fields, and we can access them anytime:

print(classifier.iterationsLimit);
print(classifier.probabilityThreshold);
// and so on

import 'package:ml_algo/ml_algo.dart';
import 'package:ml_dataframe/ml_dataframe.dart';
import 'package:ml_preprocessing/ml_preprocessing.dart';

void main() async {
  // Another option - to use a toy dataset:
  // final samples = getPimaIndiansDiabetesDataFrame();
  final samples = await fromCsv('datasets/pima_indians_diabetes_database.csv', headerExists: true);
  final targetColumnName = 'Outcome';
  final splits = splitData(samples, [0.7]);
  final validationData = splits[0];
  final testData = splits[1];
  final validator = CrossValidator.kFold(validationData, numberOfFolds: 5);
  final createClassifier = (DataFrame samples) =>
    LogisticRegressor(
      samples
      targetColumnName,
    );
  final scores = await validator.evaluate(createClassifier, MetricType.accuracy);
  final accuracy = scores.mean();
  
  print('accuracy on k fold validation: ${accuracy.toStringAsFixed(2)}');

  final testSplits = splitData(testData, [0.8]);
  final classifier = createClassifier(testSplits[0], targetNames);
  final finalScore = classifier.assess(testSplits[1], targetNames, MetricType.accuracy);
  
  print(finalScore.toStringAsFixed(2));

  await classifier.saveAsJson('diabetes_classifier.json');
}

import 'package:flutter/services.dart' show rootBundle;
import 'package:ml_algo/ml_algo.dart';
import 'package:ml_dataframe/ml_dataframe.dart';
import 'package:ml_preprocessing/ml_preprocessing.dart';

void main() async {
  final rawCsvContent = await rootBundle.loadString('assets/datasets/pima_indians_diabetes_database.csv');
  // Another option - to use a toy dataset:
  // final samples = getPimaIndiansDiabetesDataFrame();
  final samples = DataFrame.fromRawCsv(rawCsvContent);
  final targetColumnName = 'Outcome';
  final splits = splitData(samples, [0.7]);
  final validationData = splits[0];
  final testData = splits[1];
  final validator = CrossValidator.kFold(validationData, numberOfFolds: 5);
  final createClassifier = (DataFrame samples) =>
    LogisticRegressor(
      samples
      targetColumnName,
    );
  final scores = await validator.evaluate(createClassifier, MetricType.accuracy);
  final accuracy = scores.mean();
  
  print('accuracy on k fold validation: ${accuracy.toStringAsFixed(2)}');

  final testSplits = splitData(testData, [0.8]);
  final classifier = createClassifier(testSplits[0], targetNames);
  final finalScore = classifier.assess(testSplits[1], targetNames, MetricType.accuracy);
  
  print(finalScore.toStringAsFixed(2));

  await classifier.saveAsJson('diabetes_classifier.json');
}

Let's try to predict house prices using linear regression and the famous Boston Housing dataset. The dataset contains 13 independent variables and 1 dependent variable - medv which is the target one (you can find the dataset in e2e/_datasets/housing.csv).

Again, first we need to download the file and create a dataframe. The dataset is headless, we may either use autoheader or provide our own header. Let's use autoheader in our example:

Just provide a proper path to your downloaded file and use a function-factory fromCsv from ml_dataframe package to read the file:

final samples = await fromCsv('datasets/housing.csv', headerExists: false, columnDelimiter: ' ');

It's needed to add the dataset to the flutter assets by adding the following config in the pubspec.yaml:

flutter:
  assets:
    - assets/datasets/housing.csv

You need to create the assets directory in the file system and put the dataset's file there. After that you can access the dataset:

import 'package:flutter/services.dart' show rootBundle;
import 'package:ml_dataframe/ml_dataframe.dart';

final rawCsvContent = await rootBundle.loadString('assets/datasets/housing.csv');
final samples = DataFrame.fromRawCsv(rawCsvContent, fieldDelimiter: ' ');

Data in this file is represented by 505 records and 13 features. The 14th column is a target. Since we use autoheader, the target's name is autogenerated and it is col_13. Let's store it in a variable:

final targetName = 'col_13';

then let's shuffle the data:

final shuffledSamples = samples.shuffle();

Now it's the time to prepare data splits. Let's split the data into train and test subsets using the library's splitData function:

final splits = splitData(samples, [0.8]);
final trainData = splits[0];
final testData = splits[1];

splitData accepts a DataFrame instance as the first argument and ratio list as the second one. Now we have 80% of our data as a train set and 20% as a test set.

Let's train the model:

final model = LinearRegressor(trainData, targetName);

By default, LinearRegressor uses a closed-form solution to train the model. One can also use a different solution type, e.g. stochastic gradient descent algorithm:

final model = LinearRegressor.SGD(
  shuffledSamples
  targetName,
  iterationLimit: 90,
);

or linear regression based on coordinate descent with Lasso regularization:

final model = LinearRegressor.lasso(
  shuffledSamples,
  targetName,
  iterationLimit: 90,
);

Next, we should evaluate performance of our model:

final error = model.assess(testData, MetricType.mape);

print(error);

If we are fine with the error, we can save the model for the future use:

await model.saveAsJson('housing_model.json');

Later we may use our trained model for prediction:

import 'dart:io';
import 'package:ml_algo/ml_algo.dart';
import 'package:ml_dataframe/ml_dataframe.dart';

void main() async {
  final file = File('housing_model.json');
  final encodedModel = await file.readAsString();
  final model = LinearRegressor.fromJson(encodedModel);
  final unlabelledData = await fromCsv('some_unlabelled_data.csv');
  final prediction = model.predict(unlabelledData);
    
  print(prediction.header);
  print(prediction.rows);
}

import 'package:ml_algo/ml_algo.dart';
import 'package:ml_dataframe/ml_dataframe.dart';

void main() async {
  final samples = (await fromCsv('datasets/housing.csv', headerExists: false, columnDelimiter: ' ')).shuffle();
  final targetName = 'col_13';
  final splits = splitData(samples, [0.8]);
  final trainData = splits[0];
  final testData = splits[1];
  final model = LinearRegressor(trainData, targetName);
  final error = model.assess(testData, MetricType.mape);
  
  print(error);

  await classifier.saveAsJson('housing_model.json');
}

import 'package:flutter/services.dart' show rootBundle;
import 'package:ml_algo/ml_algo.dart';
import 'package:ml_dataframe/ml_dataframe.dart';

void main() async {
  final rawCsvContent = await rootBundle.loadString('assets/datasets/housing.csv');
  final samples = DataFrame.fromRawCsv(rawCsvContent, fieldDelimiter: ' ').shuffle();
  final targetName = 'col_13';
  final splits = splitData(samples, [0.8]);
  final trainData = splits[0];
  final testData = splits[1];
  final model = LinearRegressor(trainData, targetName);
  final error = model.assess(testData, MetricType.mape);
    
  print(error);
  
  await classifier.saveAsJson('housing_model.json');
}

Let's try to classify data from a well-known Iris dataset using a non-linear algorithm - decision trees

First, you need to download the data and place it in a proper place in your file system. To do so you should follow the instructions which are given in the Logistic regression section. Or you may use getIrisDataFrame function that returns ready to use DataFrame instance filled with Irisdataset.

After loading the data, it's needed to preprocess it. We should drop the Id column since the column doesn't make sense. Also, we need to encode the 'Species' column - originally, it contains 3 repeated string labels, to feed it to the classifier it's needed to convert the labels into numbers:

import 'package:ml_algo/ml_algo.dart';
import 'package:ml_dataframe/ml_dataframe.dart';
import 'package:ml_preprocessing/ml_preprocessing.dart';

void main() async {
    final samples = getIrisDataFrame()
      .shuffle()
      .dropSeries(names: ['Id']);
    
    final pipeline = Pipeline(samples, [
      toIntegerLabels(
        columnNames: ['Species'], // Here we convert strings from 'Species' column into numbers
      ),
    ]);
}

Next, let's create a model:

final model = DecisionTreeClassifier(
  processed,
  'Species',
  minError: 0.3,
  minSamplesCount: 5,
  maxDepth: 4,
);

As you can see, we specified 3 hyperparameters: minError, minSamplesCount and maxDepth. Let's look at the parameters in more detail:

• minError. A minimum error on a tree node. If the error is less than or equal to the value, the node is considered a leaf.
• minSamplesCount. A minimum number of samples on a node. If the number of samples is less than or equal to the value, the node is considered a leaf.
• maxDepth. A maximum depth of the resulting decision tree. Once the tree reaches the maxDepth, all the level's nodes are considered leaves.

All the parameters serve as stopping criteria for the tree building algorithm.

Now we have a ready to use model. As usual, we can save the model to a JSON file:

await model.saveAsJson('path/to/json/file.json');

Unlike other models, in the case of a decision tree, we can visualise the algorithm result - we can save the model as an SVG file:

await model.saveAsSvg('path/to/svg/file.svg');

Once we saved it, we can open the file through any image viewer, e.g. through a web browser. An example of the resulting SVG image:

Let's take a look at another field of machine learning - data retrieval. The field is represented by a family of algorithms, one of them is KDTree which is exposed by the library.

KDTree is an algorithm that divides the whole search space into partitions in form of the binary tree which makes it efficient to retrieve data.

Let's retrieve some data points through a kd-tree built on the Iris dataset.

First, we need to prepare the data. To do so, it's needed to load the dataset. For this purpose, we may use getIrisDataFrame function from ml_dataframe. The function returns prefilled with the Iris data DataFrame instance:

import 'package:ml_algo/ml_algo.dart';
import 'package:ml_dataframe/ml_dataframe.dart';

void main() {
  final originalData = getIrisDataFrame();
}

Since the dataset contains Id column that doesn't make sense and Species column that contains text data, we need to drop these columns:

import 'package:ml_algo/ml_algo.dart';
import 'package:ml_dataframe/ml_dataframe.dart';

void main() {
  final originalData = getIrisDataFrame();
  final data = originalData.dropSeries(names: ['Id', 'Species']);
}

Next, we can build the tree:

import 'package:ml_algo/ml_algo.dart';
import 'package:ml_dataframe/ml_dataframe.dart';

void main() {
  final originalData = getIrisDataFrame();
  final data = originalData.dropSeries(names: ['Id', 'Species']);
  final tree = KDTree(data);
}

And query nearest neighbours for an arbitrary point. Let's say, we want to find 5 nearest neighbours for the point [6.5, 3.01, 4.5, 1.5]:

import 'package:ml_algo/ml_algo.dart';
import 'package:ml_dataframe/ml_dataframe.dart';
import 'package:ml_linalg/vector.dart';

void main() {
  final originalData = getIrisDataFrame();
  final data = originalData.dropSeries(names: ['Id', 'Species']);
  final tree = KDTree(data);
  final neighbourCount = 5;
  final point = Vector.fromList([6.5, 3.01, 4.5, 1.5]);
  final neighbours = tree.query(point, neighbourCount);
 
  print(neighbours);
}

The last instruction prints the following:

(Index: 75, Distance: 0.17349341930302867), (Index: 51, Distance: 0.21470911402365767), (Index: 65, Distance: 0.26095956499211426), (Index: 86, Distance: 0.29681616124778537), (Index: 56, Distance: 0.4172527193942372))

The nearest point has an index 75 in the original data. Let's check a record at the index:

import 'package:ml_dataframe/ml_dataframe.dart';

void main() {
  final originalData = getIrisDataFrame();
 
  print(originalData.rows.elementAt(75));
}

It prints the following:

(76, 6.6, 3.0, 4.4, 1.4, Iris-versicolor)

Remember, we dropped Id and Species columns which are the very first and the very last elements in the output, so the rest elements, 6.6, 3.0, 4.4, 1.4 look quite similar to our target point - 6.5, 3.01, 4.5, 1.5, so the query result makes sense.

If you want to use KDTree outside the ml_algo ecosystem, meaning you don't want to use ml_linalg and ml_dataframe packages in your application, you may import only KDTree library and use fromIterable constructor and queryIterable method to perform the query:

import 'package:ml_algo/kd_tree.dart';

void main() async {
  final tree = KDTree.fromIterable([
    // some data here
  ]);
  final neighbourCount = 5;
  final neighbours = tree.queryIterable([/* some point here */], neighbourCount);
 
  print(neighbours);
}

As usual, we can persist our tree by saving it to a JSON file:

import 'dart:io';
import 'package:ml_algo/ml_algo.dart';
import 'package:ml_dataframe/ml_dataframe.dart';

void main() {
  final originalData = getIrisDataFrame();
  final data = originalData.dropSeries(names: ['Id', 'Species']);
  final tree = KDTree(data);
 
  // ...

  await tree.saveAsJson('path/to/json/file.json');
 
  // ...

  final file = await File('path/to/json/file.json').readAsString();
  final encodedTree = jsonDecode(file) as Map<String, dynamic>;
  final restoredTree = KDTree.fromJson(encodedTree);

  print(restoredTree);
}

Someday our previously shining model can degrade in terms of prediction accuracy - in this case, we can retrain it. Retraining means simply re-running the same learning algorithm that was used to generate our current model keeping the same hyperparameters but using a new data set with the same features:

import 'dart:io';

final fileName = 'diabetes_classifier.json';
final file = File(fileName);
final encodedModel = await file.readAsString();
final classifier = LogisticRegressor.fromJson(encodedModel);

// ... 
// here we do something and realize that our classifier performance is not so good
// ...

final newData = await fromCsv('path/to/dataset/with/new/data/to/retrain/the/classifier');
final retrainedClassifier = classifier.retrain(newData);

The workflow with other predictors (SoftmaxRegressor, DecisionTreeClassifier and so on) is quite similar to the described above for LogisticRegressor, feel free to experiment with other models.

Sometimes you may get NaN or Infinity as a value of your score, or it may be equal to some inconceivable value (extremely big or extremely low). To prevent so, you need to find a proper value of the initial learning rate, and also you may choose between the following learning rate strategies: constant, timeBased, stepBased and exponential:

final createClassifier = (DataFrame samples) =>
    LogisticRegressor(
      ...,
      initialLearningRate: 1e-5,
      learningRateType: LearningRateType.timeBased,
      ...,
    );

• Linear Regression in Dart
• Ordinary Least Squares (OLS) problem
• Closed-Form solution for OLS in Dart
• Gradient Descent in Dart

If you have questions, feel free to text me on

• X
• Telegram
• Facebook
• Linkedin