blayze

A fast and flexible Bayesian Naive Bayes implementation for the JVM written in Kotlin.

Fully supports the online learning paradigm, in which data, and even new features, are added as they become available.
Reasonably fast and memory efficient. We've trained a document classifier with tens of thousands of classes on hundreds of thousands of documents, and ironed out most of the hot-spots.
Naturally works with few samples, by integrating out the uncertainty on estimated parameters.
Models and data structures are immutable such that they are concurrency friendly.
Efficient serialization and deserialization using protobuf.
Missing and unknown features at prediction time are properly handled.
Minimal dependencies.

Usage

Get the latest artifact from maven central

//Java 9
Model model = new Model().batchAdd(List.of(new Update( //Models are immutable
        new Inputs( // Supports multiple feature types
                Map.of( //Text features
                        "subject", "Attention, is it true?", //features are named.
                        "body", "Good day dear beneficiary. This is Secretary to president of Benin republic is writing this email ..." // multiple features of the same type have different names
                ),
                Map.of( //Categorical features
                        "sender", "WWW.@galaxy.ocn.ne.jp"
                ),
                Map.of( //Gaussian features
                        "n_words", 482.
                )
        ),
        "spam" // the outcome, in this case spam.
)));

Map<String, Double> predictions = model.predict(new Inputs(/*...*/));// e.g. {"spam": 0.624, "ham": 0.376}

Built With

Kotlin - Language
Maven - Dependency Management
Protocol Buffers - Serialization

Versioning

We use SemVer for versioning.

Release a new version

Create a Sonatype account
- The created username and password will be referred to as <sonatype_user> and <sonatype_pwd, respectively.
Create a Sonatype Jira ticket of type Publishing Support requesting access to com.tradeshift.blayze. It must be approved by an existing user with write access. It can take a couple of days before access is granted.
Create a PR that updates the version in pom.xml along with the code changes. Merge it to master or v4 once it is approved.
Generate a gpg key: gpg --gen-key
List gpg keys: gpg --list-keys
Extract the key id of the previously generated gpg key. It will be referred to as <gpg_key_id> from now
Encrypt <sonatype_pwd> using mvn --encrypt-password. The encrypted value is referred to as <sonatype_pwd_enc>

Create a new server in ~/.m2/settings.xml

<settings>
<servers>
    <server>
        <id>ossrh-blayze</id>
        <username><sonatype_user></username>
        <password><sonatype_pwd_enc></password>
    </server>      
</servers>
</settings>

Run mvn clean deploy -P release -Dgpg.keyname=<gpg_key_id>
For further details, check Sonatype documentation

Backwards compatibility

We publish security updates for major version 4.x.x (branch v4) as well as 6.x.x (branch master)

Tradeshift / blayze

readme