I have loaded a number of Reuter news wire articles (1986) into Spark 2.2 and want do some topic learning using LDA
+--------------------+--------------------+----+
| title| body| id|
+--------------------+--------------------+----+
|FED SAYS IT SETS ...| |5434|
|MIM TO ACQUIRE ST...|Mount Isa Mines H...|5435|
|MAGNA <MAGAF> CRE...|Magna Internation...|5436|
|J.W. MAYS INC <MA...|Shr 2.27 dlrs vs ...|5437|
I have set up a pipeline
val pipeline = new Pipeline().setStages(Array(tokenizer, stopWordsRemover, vectorizer, lda))
run the model
val pipelineModel = pipeline.fit(corpus)
I can access the LDA (EM) model using
val ldaModel = pipelineModel.stages(3).asInstanceOf[DistributedLDAModel]
I can see the topics using
ldaModel.describeTopics(maxTermsPerTopic = 5).show()
which after a bit of DF manipulation, it gives topics and their associated terms and probabilities
+-------+----------+--------------------+
|topicId| term| probability|
+-------+----------+--------------------+
| 0| company| 0.08715003585328869|
| 0| corp| 0.03355461912220357|
| 0| group|0.024083945559541863|
| 0| unit|0.016712655949244752|
| 0| stake| 0.01314416068270042|
| 1| dlrs| 0.072961342546073|
| 1| debt| 0.02826491264713813|
...
i want to map the topic distribution back to the original documents. Back in Spark 1.6 to the get the topic distribution for the document (id=5435) above, i would do the following. But topicDistributions is no longer supported.
ldaModel.topicDistributions.filter(_._1 == 5435).collect
The (Spark ML LDA API) does list two new methods but i am unclear how to use them
final val topicConcentration: DoubleParam
and final val topicDistributionCol: Param[String]
Had anyone done this?
