Signed Graph-Based Recommendation Systems

Abstract

In the era of information and intelligence, recommendation systems play a vital role in offering high-quality recommendations to their users and are pivotal in satisfying and engaging users with personalized recommendations from a wide range of available options. Recommendation systems are based on user behavior and historical information. A graph-based recommendation system with user ratings on items can be represented as a bipartite graph, with nodes corresponding to users and items, and edges corresponding to ratings. Item-based recommendation systems use item-item similarity to suggest items to users. The item-item similarity is derived from a matrix of previous ratings of items by a typically large set of users. Various techniques have been developed, but their performances have usually only been compared to a small set of alternatives. Performance evaluation is further complicated because many datasets have items that are rated highly by many users, so recommending these generically popular items can produce an apparent high performance that is unhelpful in practice. This thesis begins by arguing for a baseline score method for top-k performance evaluation metrics that avoids the problem of generic popularity and compares the performance of a wide range of known techniques, including some that have not been applied to recommendations before, to the baseline, and to each other. It is observed that some popular techniques do not perform well, while some simpler and lesser-known techniques do. Subsequently, the use of signed rating data is focused, where users can express both positive and negative valence about items. I have applied various conventional techniques to signed data and then developed several extensions that explicitly exploit signed data. Novel techniques were developed for signed data, leading to significant improvements in top-10 recommendation scores.