java-als

A Java implementation of Alternating Least Squares (ALS). Supports:

• Batch ALS
• Stochastic Gradient Descent (SGD) ALS

API documentation available here. A live, online Jupyter notebook is available here or by clicking the above launch binder button.

installation

You can use this library as a Maven dependency by adding

<dependency>
    <groupId>org.ruivieira.ml</groupId>
    <artifactId>als</artifactId>
    <version>0.0.3</version>
</dependency>

to your Maven dependencies. You should then specify the dependency location by using one of the following methods:

bintray

Add the repository information to your pom.xml:

    <profiles>
        <profile>
            <repositories>
                <repository>
                    <snapshots>
                        <enabled>false</enabled>
                    </snapshots>
                    <id>bintray-ruivieira-maven</id>
                    <name>bintray</name>
                    <url>https://dl.bintray.com/ruivieira/maven</url>
                </repository>
                <repository>
                    <id>jcenter</id>
                    <url>https://jcenter.bintray.com/</url>
                </repository>
            </repositories>
            <pluginRepositories>
                <pluginRepository>
                    <snapshots>
                        <enabled>false</enabled>
                    </snapshots>
                    <id>bintray-ruivieira-maven</id>
                    <name>bintray-plugins</name>
                    <url>https://dl.bintray.com/ruivieira/maven</url>
                </pluginRepository>
            </pluginRepositories>
            <id>bintray</id>
        </profile>
    </profiles>

locally

Or by downloading this reposity and installing the library locally by running

$ mvn clean install

usage

batch ALS

A minimal example consists on parsing a ratings dataset into a collection such as List<Rating> ratings. Given this collection, a sparse ratings matrix can be created using

SparseRealMatrix R = ALSUtils.toMatrix(ratings);

The initial latent factors can then be created with

LatentFactors factors = LatentFactors.create(nUsers, nItems, rank);

We create a BatchALS instance

BatchALS als = new BatchALS(R, rank, alpha, beta);

and the ALS process can then be repeated for n iterations:

for (int iter = 0 ; iter < n ; iter++) {
	factors = als.run(factors);
}

After which, we can use the factors to approximate the ratings matrix, which will provide the rating predictions.

RealMatrix approximation = ALSUtils.approximate(factors);