Is there more efficient way to implement cosine similarity in PySpark 1.6?

Question

I am trying to compute cosine similarity between a given user_id from user's table and another table with movies in order to sort out most similar movies to recommend.

Cosine similarity: = dot(a,b) / (norm(a) * norm(b)) or dot(a,b)/sqrt((dot(a)*dot(b))

df = self.given_user.crossJoin(self.movies_df)
df = df.select('userId', 'movieId', 'user_features', 'movie_features')
df = df.rdd.map(lambda x: (x[0], x[1], x[2], x[3], float(np.dot(np.array(x[2]), np.array(x[3]))))).toDF(df.columns + ['dotxy'])
df = df.rdd.map(lambda x: (x[0], x[1], x[2], x[3], x[4], float(np.dot(np.array(x[2]), np.array(x[2]))))).toDF(df.columns + ['dotxx'])
df = df.rdd.map(lambda x: (x[0], x[1], x[2], x[3], x[4], x[5], float(np.dot(np.array(x[3]), np.array(x[3]))))).toDF(df.columns + ['dotyy'])
output = df.withColumn('cosine_sim', F.col("dotxy") / F.sqrt(F.col("dotxx") * F.col("dotyy")))

output.select('userId', 'movieId', 'dotxy', 'dotxx', 'dotyy', 'cosine_sim').orderBy('cosine_sim', ascending=False).show(5)

The resulting output looks like:

+------+-------+-----+-----+-----+----------+
|userId|movieId|dotxy|dotxx|dotyy|cosine_sim|
+------+-------+-----+-----+-----+----------+
|    18|   1430|  1.0|  0.5|  2.0|       1.0|
|    18|   2177|  1.0|  0.5|  2.0|       1.0|
|    18|   1565|  1.0|  0.5|  2.0|       1.0|
|    18|    415|  1.0|  0.5|  2.0|       1.0|
|    18|   1764|  1.0|  0.5|  2.0|       1.0|
+------+-------+-----+-----+-----+----------+

Is there more efficient/ compact way of cosine similarity function implementation in PySpark 1.6?

Answer 1

You could use more numpy functions.

import numpy as np

df = spark.createDataFrame([(18, 1, [1, 0, 1], [1, 1, 1])]).toDF('userId','movieId','user_features','movie_features')

df.rdd.map(lambda x: (x[0], x[1], x[2], x[3], float(np.dot(np.array(x[2]), np.array(x[3])) / (np.linalg.norm(np.array(x[2])) * np.linalg.norm(np.array(x[3])))))).toDF(df.columns + ['cosine_sim']).show()

+------+-------+-------------+--------------+------------------+
|userId|movieId|user_features|movie_features|       cosine_sim |
+------+-------+-------------+--------------+------------------+
|    18|      1|    [1, 0, 1]|     [1, 1, 1]|0.8164965809277259|
+------+-------+-------------+--------------+------------------+