This package implements coordinate-ascent variational inference for sparse exchangeable bipartite graph model described in the article Exchangeable Modelling of Relational Data: Checking Sparsity, Data Splitting, and Sparse Exchangeable Poisson Matrix Factorization. This works well out of the box for discovering latent structure in large matrices with entries that are natural numbers. The algorithm is well-adapted to handling matrices where many entries are 0. Inference generally runs quickly on GPU.

This roughly corresponds to the Poisson matrix factorization model of Scalable Recommendation with Poisson Factorization in the special case that sparsity parameters su and si are both negative.

• Python in [2.7.x]
• Tensorflow in [+1.1.x]
• Tensorflow w/ GPU in [+1.1.x] for GPU mode
• 64-bit architecture

We provide two scripts:

Simulates a dataset with given parameter setting; Splits the data into Train, Lookup and Holdout set

For usage, run:

cd scripts
python simulate_data.py -h

Runs the model on a given dataset.

Requires three files in the data directory: train.pkl, test_lookup.pkl, test_holdout.pkl For the format in which data should be stored, check: scripts/simulate_data.py

For usage, run:

cd scripts
python run.py -h

• Ekansh Sharma
• Victor Veitch

• Veitch, Sharma, Naulet, and Roy, Exchangeable Modelling of Relational Data: Checking Sparsity, Data Splitting, and Sparse Exchangeable Poisson Matrix Factorization, 2017
• Naulet, Sharma, Veitch, and Roy, An estimator for the tail-index of graphex processes, 2017