org.nield.kotlinstatistics.NaiveBayesClassifier.kt Maven / Gradle / Ivy

Go to download
Show more of this group Show more artifacts with this name
Show all versions of kotlin-statistics Show documentation
Statistical and analytical extensions for Kotlin
There is a newer version: 1.2.1
package org.nield.kotlinstatistics

import java.util.concurrent.atomic.AtomicBoolean
import kotlin.math.ln

/**
 * Returns a `NaiveBayesClassifier` that associates each set of `F` features from an item `T` with a category `C`.
 * New sets of features `F` can then be used to predict a category `C`.
 */
fun  Iterable.toNaiveBayesClassifier(
        featuresSelector: ((T) -> Iterable),
        categorySelector: ((T) -> C),
        observationLimit: Int = Int.MAX_VALUE,
        k1: Double = 0.5,
        k2: Double = k1 * 2.0
) = NaiveBayesClassifier(observationLimit, k1,k2).also { nbc ->

    forEach { nbc.addObservation(categorySelector(it), featuresSelector(it)) }
}

/**
 * Returns a `NaiveBayesClassifier` that associates each set of `F` features from an item `T` with a category `C`.
 * New sets of features `F` can then be used to predict a category `C`.
 */
fun  Sequence.toNaiveBayesClassifier(
        featuresSelector: ((T) -> Iterable),
        categorySelector: ((T) -> C),
        observationLimit: Int = Int.MAX_VALUE,
        k1: Double = 0.5,
        k2: Double = k1 * 2.0
) = NaiveBayesClassifier(observationLimit, k1,k2).also { nbc ->
    forEach { nbc.addObservation(categorySelector(it), featuresSelector(it)) }
}

/**
 * A `NaiveBayesClassifier` that associates each set of `F` features from an item `T` with a category `C`.
 * New sets of features `F` can then be used to predict a category `C`.
 */
class NaiveBayesClassifier(
        val observationLimit: Int = Int.MAX_VALUE,
        val k1: Double = 0.5,
        val k2: Double = k1 * 2.0

) {

    private val _population = mutableListOf>()
    private val modelStale = AtomicBoolean(false)

    @Volatile
    private var probabilities = mapOf, FeatureProbability>()
    val population: List> get() = _population

    /**
     * Adds an observation of features to a category
     */
    fun addObservation(category: C, features: Iterable) {
        if (_population.size == observationLimit) {
            _population.removeAt(0)
        }
        _population += BayesInput(category, features.toSet())
        modelStale.set(true)
    }

    /**
     * Adds an observation of features to a category
     */
    fun addObservation(category: C, vararg features: F) = addObservation(category, features.asList())

    private fun rebuildModel() {

        probabilities = _population.asSequence().flatMap { it.features.asSequence() }
                .distinct()
                .flatMap { f ->
                    _population.asSequence()
                            .map { it.category }
                            .distinct()
                            .map { c -> FeatureProbability.Key(f,c) }
                }.map { it to FeatureProbability(it.feature, it.category, this) }
                .toMap()

        modelStale.set(false)
    }

    /**
     * Returns the categories that have been captured by the model so far.
     */
    val categories get() = probabilities.keys.asSequence().map { it.category }.toSet()

    /**
     * Predicts a category `C` for a given set of `F` features
    */
    fun predict(vararg features: F) = predictWithProbability(features.toSet())?.category

    /**
     * Predicts a category `C` for a given set of `F` features
     */
    fun predict(features: Iterable) = predictWithProbability(features)?.category

    /**
     * Predicts a category `C` for a given set of `F` features, but also returns the probability of that category being correct.
     */
    fun predictWithProbability(vararg features: F) = predictWithProbability(features.toSet())

    /**
     * Predicts a category `C` for a given set of `F` features, but also returns the probability of that category being correct.
     */
    fun predictWithProbability(features: Iterable): CategoryProbability? {
        if (modelStale.get()) rebuildModel()

        val f = features.toSet()

        return categories.asSequence()
                .filter { c ->  population.any { it.category == c} && probabilities.values.any { it.feature in f } }
                .map { c ->
                    val probIfCategory = probabilities.values.asSequence().filter { it.category == c }.map {
                        if (it.feature in f) {
                            ln(it.probability)
                        } else {
                            ln(1.0 - it.probability)
                        }
                    }.sum().let(Math::exp)

                    val probIfNotCategory = probabilities.values.asSequence().filter { it.category == c }.map {
                        if (it.feature in f) {
                            ln(it.notProbability)
                        } else {
                            ln(1.0 - it.notProbability)
                        }
                    }.sum().let(Math::exp)

                    CategoryProbability(category = c, probability = probIfCategory / (probIfCategory + probIfNotCategory))
                }.filter { it.probability >= .1 }
                .sortedByDescending { it.probability }
                .firstOrNull()

    }

    class FeatureProbability(val feature: F, val category: C, nbc: NaiveBayesClassifier) {

        val probability = (nbc.k1 + nbc.population.count { it.category == category && feature in it.features } ) /
                (nbc.k2 + nbc.population.count { it.category == category })

        val notProbability = (nbc.k1 + nbc.population.count { it.category != category && feature in it.features } ) /
                (nbc.k2 + nbc.population.count { it.category != category })

        data class Key(val feature: F, val category: C)
        val key get() = Key(feature, category)

    }
}

data class BayesInput(val category: C, val features: Set)

data class CategoryProbability(val category: C, val probability: Double)