Bayesian Personalized Ranking using Matrix Factorization

A recomender model used to recommend items to a user based on his purchase history. ALthough there exists models based on Matrix Factorization used for item reccommendation/prediction , BPR incorporates item ranking into the optimization criterion and training data points . Furthermore the decomposed matrix is learned using a stochastic gradient descent approach using bootstrap sampling which proves to work better than standard gradient descent.

This implementation is based on the following paper : Rendle, Steffen, et al. "BPR: Bayesian personalized ranking from implicit feedback." Proceedings of the twenty-fifth conference on uncertainty in artificial intelligence. AUAI Press, 2009. https://arxiv.org/pdf/1205.2618

Getting Started

• create BPR object
• classs method train() to begin training
• class method evaluate() to evaluate model for a particular (user, item_i, item_j) triplet
• use method close() to close tf session after inferencing is done

Dataset

Online Retail Data Set:

""This is a transnational data set which contains all the transactions occurring between 01/12/2010 and 09/12/2011 for a UK-based and registered non-store online retail.The company mainly sells unique all-occasion gifts. Many customers of the company are wholesalers.""

Daqing Chen, Sai Liang Sain, and Kun Guo, Data mining for the online retail industry: A case study of RFM model-based customer segmentation using data mining, Journal of Database Marketing and Customer Strategy Management, Vol. 19, No. 3, pp. 197–208, 2012 (Published online before print: 27 August 2012. doi: 10.1057/dbm.2012.17).

Dependencies

• python 3+
• numpy
• pandas
• scipy
• tensorflow==1.9.0

Additional Notes

• Convergence metric used is AUC (Area Under the curve) on the test data. the paper describes creation of the test data and evaluation methodology in section 6.2