Machine Learning @ Amazon

Abstract

In this talk, I will first provide an overview of the key Machine Learning (ML) applications we are developing at Amazon. I will then describe a matrix factorization model that we have developed for making product recommendations âĂŞ the salient characteristics of the model are: (1) It uses a Bayesian approach to handle data sparsity, (2) It leverages user and item features to handle the cold start problem, and (3) It introduces latent variables to handle multiple personas associated with a user account (e.g. family members). Our experimental results with synthetic and real-life datasets show that leveraging user and item features, and incorporating user personas enables our model to provide lower RMSE and perplexity compared to baselines.