[SPARK-24258] SPIP: Improve PySpark support for ML Matrix and Vector types

• Type: New Feature
• Status: Resolved
• Priority: Major
• Resolution: Incomplete
• Affects Version/s: 2.3.0
• Component/s: ML, PySpark
• Labels: bulk-closed
• Assignee: Unassigned
• Reporter: Leif Mortenson
• Created: 12/May/18 22:43
• Updated: 08/Oct/19 05:42
• Resolved: 08/Oct/19 05:42

Description

Background and Motivation:

In Spark ML (pyspark.ml.linalg), there are four column types you can construct, SparseVector, DenseVector, SparseMatrix, and DenseMatrix. In PySpark, you can construct one of these vectors with VectorAssembler, and then you can run python UDFs on these columns, and use toArray() to get numpy ndarrays and do things with them. They also have a small native API where you can compute dot(), norm(), and a few other things with them (I think these are computed in scala, not python, could be wrong).

For statistical applications, having the ability to manipulate columns of matrix and vector values (from here on, I will use the term tensor to refer to arrays of arbitrary dimensionality, matrices are 2-tensors and vectors are 1-tensors) would be powerful. For example, you could use PySpark to reshape your data in parallel, assemble some matrices from your raw data, and then run some statistical computation on them using UDFs leveraging python libraries like statsmodels, numpy, tensorflow, and scikit-learn.

I propose enriching the pyspark.ml.linalg types in the following ways:

1. Expand the set of column operations one can apply to tensor columns beyond the few functions currently available on these types. Ideally, the API should aim to be as wide as the numpy ndarray API, but would wrap Breeze operations. For example, we should provide DenseVector.outerProduct() so that a user could write something like df.withColumn("XtX", df["X"].outerProduct(df["X"])).
2. Make sure all ser/de mechanisms (including Arrow) understand these types, and faithfully represent them as natural types in all languages (in scala and java, Breeze objects, in python, numpy ndarrays rather than the pyspark.ml.linalg types that wrap them, in SparkR, I'm not sure what, but something natural) when applying UDFs or collecting with toPandas().
3. Improve the construction of these types from scalar columns. The VectorAssembler API is not very ergonomic. I propose something like df.withColumn("predictors", Vector.of(df["feature1"], df["feature2"], df["feature3"])).

Target Personas:

Data scientists, machine learning practitioners, machine learning library developers.

Goals:

This would allow users to do more statistical computation in Spark natively, and would allow users to apply python statistical computation to data in Spark using UDFs.

Non-Goals:

I suppose one non-goal is to reimplement something like statsmodels using Breeze data structures and computation. That could be seen as an effort to enrich Spark ML itself, but is out of scope of this effort. This effort is just to make it possible and easy to apply existing python libraries to tensor values in parallel.

Proposed API Changes:

Add the above APIs to PySpark and the other language bindings. I think the list is:

1. pyspark.ml.linalg.Vector.of(*columns)
2. pyspark.ml.linalg.Matrix.of(<not sure what goes here, maybe we don't provide this>)
3. For each of the matrix and vector types in pyspark.ml.linalg, add more methods like outerProduct, matmul, kron, etc. https://docs.scipy.org/doc/numpy-1.14.0/reference/routines.linalg.html has a good list to look at.

Also, change python UDFs so that these tensor types are passed to the python function not as {Sparse,Dense}{Matrix,Vector} objects that wrap numpy.ndarray, but as numpy.ndarray objects by themselves, and interpret return values that are numpy.ndarray objects back into the spark types.