Recommender Systems with Sparse Data

This directory has a set of examples of how to build various kinds of recommender systems using MXNet. It also includes a set of tools for using sparse data.

Examples

The examples are driven by notebook files.

• Matrix Factorization part 1: linear and non-linear models
• Matrix Factorization part 2: overfitting and deep ResNet
• Binary classification with negative sampling
• Deep Structured Semantic Model (DSSM) for content-based recommendations

Prerequisite

The plotting functionality in the above examples requires 0.12.2 version of Bokeh package. The plotting functionality throws following error when a different Bokeh version is loaded.

ValueError: PATCH-DOC message requires at least one event

Re-usable code

These examples use and demonstrate a number of layers and other tools that can be used outside of these examples. They are all available from the recotools package.

Negative Sampling

• NegativeSamplingDataIter

Loss Layers

• CosineLoss
• CrossEntropyLoss

Sparse Data Projection layers

• SparseRandomProjection
• SparseBagOfWordProjection

Acknowledgements

Thanks to xlvector for the first Matrix Factorization example that provided the basis for these examples.

MovieLens data from GroupLens.