mahout SparseVectorsFromSequenceFiles详解（8）

最后一步，生成tf-idf

调用方法是TFIDFConverter.processTfIdf，继续以tf-vectors为输入目录

先是makePartialVectors，hadoop程序，Mapper是缺省的，Reducer是TFIDFPartialVectorReducer

  @Override
  protected void reduce(WritableComparable<?> key, Iterable<VectorWritable> values, Context context)
    throws IOException, InterruptedException {
    Iterator<VectorWritable> it = values.iterator();
    if (!it.hasNext()) {
      return;
    }
    Vector value = it.next().get();
    Iterator<Vector.Element> it1 = value.iterateNonZero();
    Vector vector = new RandomAccessSparseVector((int) featureCount, value.getNumNondefaultElements());
    while (it1.hasNext()) {
      Vector.Element e = it1.next();
      if (!dictionary.containsKey(e.index())) {
        continue;
      }
      long df = dictionary.get(e.index());
      if (maxDf > -1 && df > maxDf) {
        continue;
      }
      if (df < minDf) {
        df = minDf;
      }
      vector.setQuick(e.index(), tfidf.calculate((int) e.get(), (int) df, (int) featureCount, (int) vectorCount));
    }
    if (sequentialAccess) {
      vector = new SequentialAccessSparseVector(vector);
    }

    if (namedVector) {
      vector = new NamedVector(vector, key.toString());
    }

    VectorWritable vectorWritable = new VectorWritable(vector);
    context.write(key, vectorWritable);
  }

key是文档id，value是index组成的vectors

重点的语句是

vector.setQuick(e.index(), tfidf.calculate((int) e.get(), (int) df, (int) featureCount, (int) vectorCount));

如理解tf-idf算法，上面语句应该很容易理解 -- tf * log(n/df)，这里的算法是用luceneDefaultSimilarity计算的

public class TFIDF implements Weight {

  private Similarity sim = new DefaultSimilarity();

  public TFIDF() { }

  public TFIDF(Similarity sim) {
    this.sim = sim;
  }

  @Override
  public double calculate(int tf, int df, int length, int numDocs) {
    // ignore length
    return sim.tf(tf) * sim.idf(df, numDocs);
  }
}

最后由PartialVectorMerger.mergePartialVectors将各部分merge在一起