Electronic Thesis and Dissertation Repository

Degree

Doctor of Philosophy

Program

Computer Science

Supervisor

Dr. Charles X. Ling

Abstract

Recommendation systems have been the most emerging technology in the last decade as one of the key parts in e-commerce ecosystem. Businesses offer a wide variety of items and contents through different channels such as Internet, Smart TVs, Digital Screens, etc. The number of these items sometimes goes over millions for some businesses. Therefore, users can have trouble finding the products that they are looking for. Recommendation systems address this problem by providing powerful methods which enable users to filter through large information and product space based on their preferences. Moreover, users have different preferences. Thus, businesses can employ recommendation systems to target more audiences by addressing them with personalized content. Recent studies show a significant improvement of revenue and conversion rate for recommendation system adopters. Accuracy, scalability, comprehensibility, and data sparsity are main challenges in recommendation systems. Businesses need practical and scalable recommendation models which accurately personalize millions of items for millions of users in real-time. They also prefer comprehensible recommendations to understand how these models target their users. However, data sparsity and lack of enough data about items, users and their interests prevent personalization models to generate accurate recommendations. In Chapter 1, we first describe basic definitions in recommendation systems. We then shortly review our contributions and their importance in this thesis. Then in Chapter 2, we review the major solutions in this context. Traditional recommendation system methods usually make a rating matrix based on the observed ratings of users on items. This rating matrix is then employed in different data mining techniques to predict the unknown rating values based on the known values. In a novel solution, in Chapter 3, we capture the mean interest of the cluster of users on the cluster of items in a cluster-level rating matrix. We first cluster users and items separately based on the known ratings. In a new matrix, we then present the interest of each user clusters on each item clusters by averaging the ratings of users inside each user cluster on the items belonging to each item cluster. Then, we apply the matrix factorization method on this coarse matrix to predict the future cluster-level interests. Our final rating prediction includes an aggregation of the traditional user-item rating predictions and our cluster-level rating predictions. Generating personalized recommendation for cold-start users, or users with only few feedback, is a big challenge in recommendation systems. Employing any available information from these users in other domains is crucial to improve their recommendation accuracy. Thus, in Chapter 4, we extend our proposed clustering-based recommendation model by including the auxiliary feedback in other domains. In a new cluster-level rating matrix, we capture the cluster-level interests between the domains to reduce the sparsity of the known ratings. By factorizing this cross-domain rating matrix, we effectively utilize data from auxiliary domains to achieve better recommendations in the target domain, especially for cold-start users. In Chapter 5, we apply our proposed clustering-based recommendation system to Morphio platform used in a local digital marketing agency called Arcane inc. Morphio is an smart adaptive web platform, which is designed to help Arcane to produce smart contents and target more audiences. In Morphio, agencies can define multiple versions of content including texts, images, colors, and so on for their web pages. A personalization module then matches a version of content to each user using their profiles. Our ongoing real time experiment shows a significant improvement of user conversion employing our proposed clustering-based personalization. Finally, in Chapter 6, we present a summary and conclusions for this thesis. Parts of this thesis were submitted or published in peer-review journal and conferences including ACM Transactions on Knowledge Discovery from Data and ACM Conferences on Recommender Systems.

Share

COinS