All - I'm having the same issue as mentioned at http://comments.gmane.org/gmane.comp.apache.mahout.user/18889 on Mahout 0.9. My CVB clusters describe my corpus well; however, the mapping file generated by mahout's `rowid` seems to be wayyyyyy off.
For example, there's a very obvious cluster which has keywords like "beer, stout, pale" - the only cluster to contain these keywords. In my vectordump for the p(term | topic) this cluster is at line 217. Vector dump generated by: echo `date` ": Dumping the p(term | topic) vectors to local filesystem..." $mahout_bin/mahout vectordump -i results/cvb_results/to_out \ --dictionary results/seq2sparse_results/dictionary.file-0 \ --vectorSize $NUM_KEYWORDS -sort results/cvb_results/to_out \ -o $OUTPUT_DIR/$PTOPIC_TERM_FILE -dt sequencefile And, while the results of dumping out the p(doc | topic) group all of the documents which contain the words "beer, stout, pale" together - it dumps them into cluster number 8. The dump is created via: echo `date` ": Dumping the p(doc | topic) vectors to local filesystem..." $mahout_bin/mahout vectordump -i results/cvb_results/do_out \ -sort results/cvb_results/do_out \ -o $OUTPUT_DIR/$PDOC_TOPIC_FILE -p true -c csv -n true -u true IE: the result from the p(doc | topic) dump will result in: 123 0.001,...,0.60,... Where 123 maps to a document about "beer, stout, pale" and where 0.60 is the 9th comma separated value -- thus belonging to cluster id#8 (at zero index). However, if we look at the p(term | topic) file dumped earlier, cluster id#8 has nothing to do with this document. Additionally, I wrote a script to review all of the documents belonging to any given cluster; and, all of the documents in cluster #8 actually map to the p(term|topic) entry described by cluster #217. That is to say, these are the only documents containing the ngrams / keywords that cluster #217 shows as describing it. I can't figure out where the gap is: Is it in the rowid docIndex/matrix I have? I've tried dumping the above two files without sorting as I figured that might be rearranging the ordering of cluster probabilities in the p(doc | topic) dump, but that turned up inconclusive I believe. I would love any ideas - I've been stumped on this for a little while now. Thank you, Mo
