Improving Recommender System Diversity using Variational Autoencoders
Loading...
URL
Journal Title
Journal ISSN
Volume Title
Perustieteiden korkeakoulu |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
2022-10-17
Department
Major/Subject
Data Science
Mcode
SCI3115
Degree programme
Master's Programme in ICT Innovation
Language
en
Pages
59+12
Series
Abstract
Recommender systems have been widely adopted in many use cases to help customers find relevant items in eCommerce and content recommendation platforms. Collaborative filtering algorithms are often trained to optimize accuracy, but recent user research has shown that other system features, including diversity, are also crucial to facilitate a good user experience. In this work, we aim to assess and improve the diversity of recommendations in the context of large eCommerce platforms. This research has been done in collaboration with Bol.com, the largest e-commerce retailer in the Benelux region. Diversity of recommendations has been defined in numerous ways in the literature. However, these definitions have not been evaluated with the limitations of a real-world recommender system like scalability constraints. Therefore, we first evaluate how diversity should be measured holistically and feasibly in real-world recommender systems. Second, we achieve diversity improvements in recommender systems by using variational autoencoders. These models have previously been used in natural language tasks for improving diversity, but not in recommender systems. In this study, we have used the generative nature of variational autoencoders to generate a distribution from which we sample multiple user profiles that are used to generate recommendations with higher user and item level diversity. Through empirical analysis over benchmark and real-world datasets, we show that our approach produces recommendations that are more diverse in several ways. First, a single recommendation list for a user is more diverse. Second, recommendations generated for each user over time are more diverse. We propose a novel metric called temporal inter-list diversity to measure this effect. Third, the total number of items exposed to the users increased as well. Additionally, we have done a parameter sensitivity analysis to verify to what extent the results depend on the parameter settings and help practitioners identify how to tune the different parameters in the system to achieve the desired accuracy and diversity in recommendations. We think that the proposed method and evaluations can help improve customer satisfaction and vendor exposure in recommender systems.Description
Supervisor
Pechenizkiy, MykolaThesis advisor
Gebre, BinyamWeerts, Hilde
Keywords
variational autoencoders, recommender systems, diversity, multiple user profiles