Factorization Machines is a smart general predictor introduced by Rendle in 2010, which can capture all single and pairwise interactions in a dataset. It can be applied to any real valued feature vector and also performs well on highly sparse data. An extension on FMs, namely Field Factorization Machines, proved to be a successful method in predicting advertisement clicks in the Display Advertising Challenge on Kaggle.

I built a custom Spark implementation to use in Python and Scala. To make optimum use of parallel computing in Spark, I implemented Parallel Stochastic Gradient Descent to train the FMs. This forms an alternative to Mini-batch SGD, which is currently available in MLLib to train Logistic Regression models.

This implementation shows impressive results in terms of speed and effectivness.