What's the best way to do matrix multiplication between 2 lists of sparseVectors with DataFrame-based API in pyspark?

Question

I have 2 DataFrames with the same structure: DataFrame[id: bigint, tfidf_features: vector]

I need to multiple rows in dataframe1 with rows in dataframe2. I can use a loop and do things like: dataframe1.collect()[i]['tfidf_features'].dot(dataframe2.collect()[j]['tfidf_features']).

However, I would like to use matrix multiplication, something equivalent to: np.matmul(dataframe1_tfidf_features, dataframe2_tfidf_features.T).

Answer 1

You have two choices
1. mllib.linalg.distributed.BlockMatrix convert both dataframes to block matrices and use mulitply

bm1 = IndexedRowMatrix(df1.rdd.map(lambda x: IndexedRow(x[0], x[1]))).toBlockMatrix()
bm2 = IndexedRowMatrix(df2.rdd.map(lambda x: IndexedRow(x[0], x[1]))).toBlockMatrix()
bm_result = bm1.multiply(bm2)  

2. pyspark.sql.dataframe.crossJoin crossjoin both dataframes and calculate individual element of resultant matrix and then use collect_list & sort

arr = np.array
df =df1.crossJoin(df2.select(col("id").alias("id2"),
                                 col("features").alias("features2"))

udf_mult = udf(lambda x,y = float(arr(x).dot(arr(y).T).sum()), DoubleType()) 
df = df.withColumn("val", udf_mult("features","features2")).
                         drop("features","features2")
st = struct(["id2","val"]).alias("map")
df = df.select("id", st).groupBy("id").agg(collect_list("map").alias("list"))

def sort(x):

    x = sorted(x, key=lambda x:x[0])
    y = list(map(lambda a:a[1], x))
    return(y)
udf_sort = udf(sort, ArrayType(DoubleType()))
df = df.withColumn("list", udf_sort("list"))