Attention-based mixture density recurrent networks for history-based recommendation

Abstract

Recommendation system has been widely used in search, online advertising, e-Commerce, etc. Most products and services can be formulated as a personalized recommendation problem. Based on users' past behavior, the goal of personalized history-based recommendation is to dynamically predict the user's propensity (online purchase, click, etc.) distribution over time given a sequence of previous activities. In this paper, with an e-Commerce use case, we present a novel and general recommendation approach that uses a recurrent network to summarize the history of users' past purchases, with a continuous vectors representing items, and an attention-based recurrent mixture density network, which outputs each mixture component dynamically, to accurate model the predictive distribution of future purchase. We evaluate the proposed approach on two publicly available datasets, MovieLens-20M and RecSys15. Both experiments show that the proposed approach, which explicitly models the multi-modal nature of the predictive distribution, is able to greatly improve the performance over various baselines in terms of precision, recall and nDCG. The new modeling framework proposed can be easily adopted to many domain-specific problems, such as item recommendation in e-Commerce, ads targeting in online advertising, click-through-rate modeling, etc.