Top-N Recommendations Using Deep Learning

Abstract

Recommender systems RS(s) are the smart systems which can compute the preferences of items of various users. It suggests only the most relevant and appropriate products/ information in which a particular user is interested through personalized recommendation. It plays a vital role in various social networking applications, e-commerce services and recommending products to people such as movies, garments and research papers. Collaborative filtering (CF) is a prevalent recommendation algorithm that formulates its results on the ratings or activities of users. The ultimate hypothesis backing this methodology is that users’ views can be shortlisted and aggregated in a way so as to deduce a rational prediction of the active user’s preference. CF is categorized into user oriented and item oriented techniques. User oriented methodology measures the relationship among objective users and other users, whereas item oriented methodology measures the likeness among the items that objective users’ rates / interacts with and other items. Item-based Collaborative Filtering (ICF) has been deployed in different platforms in the industry involving recommender frameworks to promote personalization as per each user’s requirements. ICF fundamentally recommends items based on the user’s profile history, maintained using different algorithms. Existing methods offer conventional solutions mostly based on linear correlation, but the accuracy and efficiency of such systems can be considerably increased by implementing custom relevant Machine Learning (ML) algorithms to cater for the complications arising due to irregular user decision making elements. The paper proposes an advanced alternative to conventional RS(s) by incorporating the Long Short Term Memory (LSTM) and Probabilistic Matrix Factorization (PMF) models to get an improved accuracy, efficiency and relevance in the recommendation results of such systems, as compared to existing solutions. Finally, RMSE is used as predictive accuracy and decision support metrics are used to compare the results on real world dataset, i.e., movieLens100K, movieLens1M and HetRec2011.