Exploring New Frontiers in Modern Recommendation Systems: GNN-based Models, Ensemble Approaches, and Conversational Interfaces
File version
Author(s)
Primary Supervisor
Wang, Can
Other Supervisors
Liew, Wee-Chung
Editor(s)
Date
Size
File type(s)
Location
License
Abstract
Recommendation systems have become an integral part of most modern web applications, greatly reducing the burden of information overload for users by providing intelligent recommendations. Typically, recommendation systems use the user's past interaction history as input to create appropriate representations. These representations are then fed into the prediction model to generate prediction scores and recommendation outcomes, which are ultimately delivered to front-end web applications. Collaborative filtering-based (CF) [1] approaches and content-based (CB) approaches [2] are two primary categories of traditional recommendation methods. With the rapid development of deep learning techniques, many new recommendation models have been proposed in recent years. These models have achieved significant improvements in terms of accuracy comparing to traditional methods. Graph Neural Networks (GNNs) [3], which are a class of neural networks that can learn representations of graph-structured data, have been widely used in recommendation systems. Besides, we are also seeing the emergence of new recommendation paradigms, such as Conversational Recommendation Systems (CRS), which engage users in interactive dialogues to refine their preferences and provide more relevant suggestions. This thesis explores modern advances in recommendation systems includes several works, focusing on three aspects: Graph Neural Networks (GNNs), recommender ensembles, and Conversational Recommendation Systems (CRS). Recommender system have become an essential component of e-commerce and entertainment platforms which provide users with personalized suggestions for items and content based on their preferences and behavior. We first examine the use of GNNs for enhancing user and item representations in recommendation systems. By exploiting the graph structure of user-item interactions, GNN-based models can capture complex relationships and dependencies between users and items, leading to improved recommendation accuracy. In this thesis, we propose two sequential recommendation models that leverage the power of GNNs for item embedding enhancement. then we investigate recommender ensembles, which combine the outputs of multiple recommendation models to generate more accurate and robust recommendations. We propose three ensemble-based approaches that explore the latest deep learning techniques, such as contrastive learning and meta learning, to improve recommendation performance. we also show our latest work on CRS, which engage users in interactive dialogues to refine their preferences and provide more relevant suggestions. We present our proposed KERL system, which incorporates information from knowledge graph embeded by pre-trained language model to adapt to user feedback and improve over time. Through a comprehensive examination of these modern techniques, this thesis contributes to the advancement of recommendation systems research and demonstrates the potential for even more effective and personalized recommendations in the future.
Journal Title
Conference Title
Book Title
Edition
Volume
Issue
Thesis Type
Thesis (PhD Doctorate)
Degree Program
Doctor of Philosophy (PhD)
School
School of Info & Comm Tech
Publisher link
Patent number
Funder(s)
Grant identifier(s)
Rights Statement
Rights Statement
The author owns the copyright in this thesis, unless stated otherwise.
Item Access Status
Note
Access the data
Related item(s)
Subject
recommendation system
deep learning
graph neural network
data mining