Link recommendation beyond homophily and an extension to recommender systems
Date
2022
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Online social media has gained soaring popularity and generated unprecedented commercial values because of their penetration into people's daily lives. As one of the essential functions of online social media, recommendation algorithms have attracted considerable attention from academia and industry. A recommendation algorithm with high prediction accuracy benefits both users and business operators significantly. Aiming at improving the prediction accuracy of the state-of-the art recommendation algorithms, this dissertation proposes three novel recommendation problems and develops novel solution methods to these problems. ☐ In the first study of this dissertation, we propose a novel link recommendation problem, and develop its solution method, namely the Diversity Preference-Aware Link Recommendation (DPA-LR) method. Traditional link recommendation methods overlook the importance of users' diversity preferences. According to social and psychology theories, people's diversity preference is an essential driver in friendship formation and should be utilized to increase link prediction accuracy. To address this research gap, we formulate the new DPA-LR problem and design its novel solution method. We then demonstrate the superiority of the DPA-LR method over benchmarks with two large-scale social network datasets. ☐ In the second study, we extend the DPA-LR method to recommender systems. Diversity and novelty are essential objectives in recommender systems to improve stakeholders' benefits by reducing user's discovery efforts and improving business operators' sales and revenue. Existing diversity and novelty-based methods indifferently increase diversity or novelty for every user, which inevitably induces the trade-off dilemma between relevance and accuracy. Moreover, different users have different preferences for recommendation diversity and novelty. Such preference should be considered by a recommendation algorithm, thereby avoiding the trade-off dilemma and increasing the prediction accuracy. ☐ To address this research gap, we propose a new Diversity and Serendipity-Aware Recommender System (DSPA-RS) problem and its solution method. Using the MovieLens-2k data set, we demonstrate the superior predictive power of the DSPA-RS method over benchmarks. ☐ In the third study, we solve the link recommendation problem for online social networks from a brand new perspective. Traditional link recommendation methods are solely homophily-based and built upon the basic assumption that people tend to befriend others similar to themselves. Recent sociology studies, however, tell a different story. Different people have different preferences in making friends, e.g., some prefer friends similar to themselves but others prefer friends with diverse backgrounds. Meanwhile, such preference is dynamic and driven by personal internal needs and social influence. Therefore, we should design the link recommendation algorithm from a new perspective of dynamic preference instead of homophily. Accordingly, we propose a novel Dynamic Preference-Based Link Recommendation problem (DyPB-LR) and a novel dynamic graph neural network framework to solve the problem.
Description
Keywords
Optimization, Recommendation algorithm, Online social networks