org.neo4j.gds.embeddings.hashgnn.HashGNN Maven / Gradle / Ivy

Go to download
/*
 * Copyright (c) "Neo4j"
 * Neo4j Sweden AB [http://neo4j.com]
 *
 * This file is part of Neo4j.
 *
 * Neo4j is free software: you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program.  If not, see .
 */
package org.neo4j.gds.embeddings.hashgnn;

import org.apache.commons.lang3.mutable.MutableLong;
import org.neo4j.gds.Algorithm;
import org.neo4j.gds.api.Graph;
import org.neo4j.gds.api.properties.nodes.NodePropertyValues;
import org.neo4j.gds.api.schema.GraphSchema;
import org.neo4j.gds.collections.ha.HugeObjectArray;
import org.neo4j.gds.core.concurrency.Concurrency;
import org.neo4j.gds.core.utils.paged.HugeAtomicBitSet;
import org.neo4j.gds.core.utils.partition.Partition;
import org.neo4j.gds.core.utils.partition.PartitionUtils;
import org.neo4j.gds.core.utils.progress.tasks.ProgressTracker;
import org.neo4j.gds.termination.TerminationFlag;

import java.util.List;
import java.util.Optional;
import java.util.Set;
import java.util.SplittableRandom;
import java.util.function.Function;
import java.util.stream.Collectors;

import static org.neo4j.gds.utils.StringFormatting.formatWithLocale;

/**
 * Based on the paper "Hashing-Accelerated Graph Neural Networks for Link Prediction"
 */
public class HashGNN extends Algorithm {
    private static final long DEGREE_PARTITIONS_PER_THREAD = 4;
    private final long randomSeed;
    private final Graph graph;
    private final SplittableRandom rng;
    private final HashGNNParameters parameters;
    private final Concurrency concurrency;
    private final MutableLong currentTotalFeatureCount = new MutableLong();

    public HashGNN(Graph graph, HashGNNParameters parameters, ProgressTracker progressTracker, TerminationFlag terminationFlag) {
        super(progressTracker);
        this.graph = graph;
        this.parameters = parameters;
        this.concurrency = parameters.concurrency();

        long tempRandomSeed = this.parameters.randomSeed().orElse((new SplittableRandom().nextLong()));
        this.randomSeed = new SplittableRandom(tempRandomSeed).nextLong();
        this.rng = new SplittableRandom(randomSeed);

        this.terminationFlag = terminationFlag;
    }

    @Override
    public HashGNNResult compute() {
        progressTracker.beginSubTask("HashGNN");

        // Since degree only very approximately reflect the min hash task workload per node we decrease the partition sizes.
        int decreasedConcurrency = Math.toIntExact(Math.min(concurrency.value() * DEGREE_PARTITIONS_PER_THREAD, graph.nodeCount()));

        var degreePartition = PartitionUtils.degreePartition(
            graph,
            new Concurrency(decreasedConcurrency),
            Function.identity(),
            Optional.of(1)
        );
        var rangePartition = PartitionUtils.rangePartition(
            concurrency,
            graph.nodeCount(),
            Function.identity(),
            Optional.of(1)
        );

        Graph graphCopy = graph.concurrentCopy();
        GraphSchema schema = graph.schema();
        List graphs = parameters.heterogeneous()
            ? schema.relationshipSchema().availableTypes()
            .stream()
            .map(rt -> graph.relationshipTypeFilteredGraph(Set.of(rt)))
            .collect(Collectors.toList())
            : List.of(graphCopy);

        var embeddingsB = constructInputEmbeddings(rangePartition);
        int embeddingDimension = (int) embeddingsB.get(0).size();

        double avgInputActiveFeatures = currentTotalFeatureCount.doubleValue() / graph.nodeCount();
        progressTracker.logInfo(formatWithLocale(
            "Density (number of active features) of binary input features is %.4f.",
            avgInputActiveFeatures
        ));

        var embeddingsA = HugeObjectArray.newArray(HugeAtomicBitSet.class, graph.nodeCount());
        embeddingsA.setAll(unused -> HugeAtomicBitSet.create(embeddingDimension));

        double avgDegree = graph.relationshipCount() / (double) graph.nodeCount();
        double upperBoundNeighborExpectedBits = embeddingDimension == 0
            ? 1
            : embeddingDimension * (1 - Math.pow(
                1 - (1.0 / embeddingDimension),
                avgDegree
            ));

        progressTracker.beginSubTask("Propagate embeddings");

        for (int iteration = 0; iteration < parameters.iterations(); iteration++) {
            terminationFlag.assertRunning();

            var currentEmbeddings = iteration % 2 == 0 ? embeddingsA : embeddingsB;
            var previousEmbeddings = iteration % 2 == 0 ? embeddingsB : embeddingsA;
            for (long i = 0; i < currentEmbeddings.size(); i++) {
                currentEmbeddings.get(i).clear();
            }

            double scaledNeighborInfluence = graph.relationshipCount() == 0 ? 1.0 : (currentTotalFeatureCount.doubleValue() / graph.nodeCount()) * parameters.neighborInfluence() / upperBoundNeighborExpectedBits;
            currentTotalFeatureCount.setValue(0);

            var hashes = HashTask.compute(
                embeddingDimension,
                scaledNeighborInfluence,
                graphs.size(),
                concurrency,
                parameters.embeddingDensity(),
                randomSeed + parameters.embeddingDensity() * iteration,
                terminationFlag,
                progressTracker
            );

            MinHashTask.compute(
                degreePartition,
                graphs,
                concurrency,
                parameters.embeddingDensity(),
                embeddingDimension,
                currentEmbeddings,
                previousEmbeddings,
                hashes,
                progressTracker,
                terminationFlag,
                currentTotalFeatureCount
            );

            double avgActiveFeatures = currentTotalFeatureCount.doubleValue() / graph.nodeCount();
            progressTracker.logInfo(formatWithLocale(
                "After iteration %d average node embedding density (number of active features) is %.4f.",
                iteration,
                avgActiveFeatures
            ));
        }

        progressTracker.endSubTask("Propagate embeddings");

        var binaryOutputVectors = (parameters.iterations() - 1) % 2 == 0 ? embeddingsA : embeddingsB;

        NodePropertyValues outputVectors = parameters.outputDimension().map(it -> {
            var denseVectors = DensifyTask.compute(
                graph,
                rangePartition,
                concurrency,
                it,
                rng,
                binaryOutputVectors,
                progressTracker,
                terminationFlag
            );
            return (NodePropertyValues) EmbeddingsToNodePropertyValues.fromDense(denseVectors);
        }).orElseGet(() -> EmbeddingsToNodePropertyValues.fromBinary(binaryOutputVectors, embeddingDimension));

        progressTracker.endSubTask("HashGNN");

        return new HashGNNResult(outputVectors);
    }

    private HugeObjectArray constructInputEmbeddings(List partition) {
        // User input parsing proves that if FeatureProperties is empty
        // then GenerateFeatures is not
        if (parameters.featureProperties().isEmpty()) {
            return GenerateFeaturesTask.compute(
                parameters.generateFeatures().get(),
                graph,
                partition,
                concurrency,
                randomSeed,
                progressTracker,
                terminationFlag,
                currentTotalFeatureCount
            );
        }
        return parameters.binarizeFeatures().map(it ->
            BinarizeTask.compute(
                graph,
                partition,
                concurrency,
                parameters.featureProperties(),
                it,
                rng,
                progressTracker,
                terminationFlag,
                currentTotalFeatureCount
            )
        ).orElseGet(() ->
            RawFeaturesTask.compute(
                concurrency,
                parameters.featureProperties(),
                progressTracker,
                graph,
                partition,
                terminationFlag,
                currentTotalFeatureCount
            )
        );
    }

    static final class MinAndArgmin {
        public int min;
        public int argMin;

        MinAndArgmin() {
            this.min = -1;
            this.argMin = Integer.MAX_VALUE;
        }
    }
}