Preparing data for LDA in spark

Question

I'm working on implementing a Spark LDA model (via the Scala API), and am having trouble with the necessary formatting steps for my data. My raw data (stored in a text file) is in the following format, essentially a list of tokens and the documents they correspond to. A simplified example:

doc XXXXX   term    XXXXX
1   x       'a'     x
1   x       'a'     x
1   x       'b'     x
2   x       'b'     x
2   x       'd'     x
...

Where the XXXXX columns are garbage data I don't care about. I realize this is an atypical way of storing corpus data, but it's what I have. As is I hope is clear from the example, there's one line per token in the raw data (so if a given term appears 5 times in a document, that corresponds to 5 lines of text).

In any case, I need to format this data as sparse term-frequency vectors for running a Spark LDA model, but am unfamiliar with Scala so having some trouble.

I start with:

import org.apache.spark.mllib.clustering.{LDA, DistributedLDAModel}
import org.apache.spark.mllib.linalg.{Vector, Vectors}
import org.apache.spark.rdd.RDD

val corpus:RDD[Array[String]] = sc.textFile("path/to/data")
    .map(_.split('\t')).map(x => Array(x(0),x(2)))

And then I get the vocabulary data I'll need to generate the sparse vectors:

val vocab: RDD[String] = corpus.map(_(1)).distinct()
val vocabMap: Map[String, Int] = vocab.collect().zipWithIndex.toMap

What I don't know is the proper mapping function to use here such that I end up with a sparse term frequency vector for each document that I can then feed into the LDA model. I think I need something along these lines...

val documents: RDD[(Long, Vector)] = corpus.groupBy(_(0)).zipWithIndex
    .map(x =>(x._2,Vectors.sparse(vocabMap.size, ???)))

At which point I can run the actual LDA:

val lda = new LDA().setK(n_topics)
val ldaModel = lda.run(documents)

Basically, I don't what function to apply to each group so that I can feed term frequency data (presumably as a map?) into a sparse vector. In other words, how do I fill in the ??? in the code snippet above to achieve the desired effect?

Answer 1

One way to handle this:

• make sure that spark-csv package is available

• load data into DataFrame and select columns of interest

  val df = sqlContext.read
      .format("com.databricks.spark.csv")
      .option("header", "true")
      .option("inferSchema", "true") // Optional, providing schema is prefered
      .option("delimiter", "\t")
      .load("foo.csv")
      .select($"doc".cast("long").alias("doc"), $"term")
  

• index term column:

  import org.apache.spark.ml.feature.StringIndexer
  
  val indexer = new StringIndexer()
    .setInputCol("term")
    .setOutputCol("termIndexed")
  
  val indexed = indexer.fit(df)
    .transform(df)
    .drop("term")
    .withColumn("termIndexed", $"termIndexed".cast("integer"))
    .groupBy($"doc", $"termIndexed")
    .agg(count(lit(1)).alias("cnt").cast("double"))
  

• convert to PairwiseRDD

  import org.apache.spark.sql.Row
  
  val pairs = indexed.map{case Row(doc: Long, term: Int, cnt: Double) => 
    (doc, (term, cnt))}
  

• group by doc:

  val docs = pairs.groupByKey
  

• create feature vectors

  import org.apache.spark.mllib.linalg.Vectors
  import org.apache.spark.sql.functions.max
  
  val n = indexed.select(max($"termIndexed")).first.getInt(0) + 1
  
  val docsWithFeatures = docs.mapValues(vs => Vectors.sparse(n, vs.toSeq))
  

• now you have all you need to create LabeledPoints or apply additional processing