Mahout cheat-sheet

clusterdump.sh

mahout clusterdump \
  -dt sequencefile \                            # format: {Integer => String}
  -d reuters-vectors/dictionary.file-* \        # dictionary: {id => word}
  -i reuters-kmeans-clusters/clusters-3-final \  # input
  -o clusters.txt \                             # output (local filesystem)
  -b 10 \                                       # format length
  -n 10                                         # number of top terms to print
  --distanceMeasure org.apache.mahout.common.distance.CosineDistanceMeasure # default is euclidean distance

kmeans.sh

mahout kmeans \
  -i ".../tfidf-vectors" \  # input vectors (output of seq2sparse)
  -o ".../kmeans-clusters" \  # output dir for clusters
  -c ".../kmeans-centroids" \  # output/input dir for centroids
  --numClusters 500 \
  --distanceMeasure org.apache.mahout.common.distance.CosineDistanceMeasure \
  --maxIter 100 \ 
  --convergenceDelta 1.0 \  # convergence threshold
  -ow \  # overwrite output folder
  -cl \  # cluster after finishing 

seq2sparse.sh

# Sequence file vectorization
mahout seq2sparse \
  -i reuters-seqfiles \  # input sequence file
  -o reuters-normalized-bigram \  # output dir
  -ow \  # overwrite output dir 
  -a org.apache.lucene.analysis.WhitespaceAnalyzer \  # analyzer (tokenizer)
  -chunk 200 \  # chunk size (MB)
  -wt tfidf \   # weighting scheme
  -s 5 \        # minimum support
  -md 3 \       # minimum document frequency
  -x 90 \       # maximum document frequency percentage
  -ng 2 \       # ngram size
  -ml 50 \      # minimum log likelihood ratio
  -seq \        # create sequential access sparse vectors (good for kmeans)
  -n 2 \        # normalization - use 2-norm (aka Euclidean norm)
  -nv \         # output named vectors, with this you'll be able to identify each document
                # in clusteredPoints folder after clustering 

seqdumper.sh

# human readable sequence file output
mahout seqdumper \
  -i /research/corpora-clustering/kmeans-clusters/clusteredPoints/part-m-00000