454 0

Novel Approaches to One-Class Collaborative Filtering by Exploiting Both Positive and Negative Preferences of Users

Novel Approaches to One-Class Collaborative Filtering by Exploiting Both Positive and Negative Preferences of Users
Alternative Author(s)
Issue Date
2021. 2
When users' preferences are shown by their implicit feedback in a one-class setting such as clicks and bookmarks (as opposed to a multi-class setting), the recommendation problem becomes harder because one-class setting has inherently less information to capture a user's taste than multi-class setting (e.g., 1-5 star rating). The One-Class Collaborative Filtering (OCCF) is a popular CF approach for such one-class settings. Existing OCCF methods have often focused on how to treat the unrated items by using only rated items, but suffer from low accuracy in dealing with sparse data. In this dissertation, we investigate how to address the shortcomings of the popular OCCF methods in handling challenging ``sparse'' dataset in one-class setting. Towards this goal, this dissertation proposes four novel OCCF approaches. First, we design a graph-theoretic OCCF approach, named as gOCCF, that exploits uninteresting items of users. Second, we propose an OCCF approach with multi-type pair-wise preferences, named as M-BPR, that exploits both unknown and uninteresting items of users. Third, we design an adversarial OCCF approach, named as ASiNE, that analyzes both positive and negative preferences based on adversarial learning. Lastly, we propose a domain-specific OCCF approach for TV show domain using a concept of `watchable interval'. First, gOCCF exploits both positive preferences (i.e., interesting items) from rated items as well as negative preferences (i.e., uninteresting items) derived from unrated items. In capturing both positive and negative preferences as a bipartite graph, further, we apply the graph shattering theory to determine the right amount of negative preferences to use. Then, we develop a suite of novel graph-based OCCF approaches based on the random walk with restart and belief propagation methods. Second, M-BPR defines fine-grained multi-type pair-wise preferences (PPs). Then, we propose a novel pair-wise approach which exploits multi-type PPs together in learning users' more detailed preferences. Third, we propose an advanced OCCF approach, named (ASiNE), represents each user of a given signed network including users' positive and negative preferences as a low-dimensional vector based on the adversarial learning. Lastly, for the domain-specific OCCF approach on TV show domain, we define a user's watchable interval which is the most important and novel concept in understanding users' true preferences. In order to reflect this new concept into the TV show recommendation, we propose a novel framework based on collaborative filtering. We validate the effectiveness of our approaches via extensive experiments using several real-life datasets such as MovieLens 100K, CiteULike, Last.fm, and Amazon. Experimental results show that our approaches effectively address the sparsity challenge and significantly outperform all of competing OCCF methods in terms of recommendation accuracy. More specifically, gOCCF and M-BPR significantly improve normalized discounted cumulative gain(nDCG)@5 by up to 48% and 80% over the existing OCCF methods, respectively. Also, our ASiNE consistently and significantly outperforms all the state-of-the-art signed NE methods in all datasets and with all metrics. Lastly, our TV show recommendation approach shows dramatic improvements in recommendation accuracy over the state-of-the-art TV show recommendation methods, outperforming them up to 429% in terms of nDCG@5.
Appears in Collections:
GRADUATE SCHOOL[S](대학원) > COMPUTER SCIENCE(컴퓨터·소프트웨어학과) > Theses (Ph.D.)
Files in This Item:
There are no files associated with this item.
RIS (EndNote)
XLS (Excel)


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.