Transferable and curious AI in sequential recommendation

• Main Author: Zhang, Yinan
• Other Authors: Miao Chun Yan
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/164474
• Institution: Nanyang Technological University

With the rapid development of E-commerce, recommendation systems have attracted more research attention in recent years, which help users make effective decisions among various items. Traditional recommendation methods usually utilize the whole user-item interaction history to model users' long-term preferences while ignoring the sequence of user behaviors. In other words, these methods assign the same importance to all interactive data, regardless of whether the interaction is completed early or recently. However, users' preferences are not static and may change overtime with the growth of age and experience. Besides, user behaviors are often sequentially dependent, such as the purchase of car insurance after the purchase of a car. Simply capturing users' long-term preferences is inadequate to accurately predict user's latest desired items. Sequential recommendation systems (SRS) are then proposed to address the above challenges, which consider the sequential dependencies when modeling users' recent preferences. More specifically, SRS take a sequence of interaction data as input and aim to predict the users' behaviors in the near future. These methods, however, usually focus on capturing users' recent interests and pay little attention to the early interactions, which often result in poor recommendation performances because users' decision making is complex and may be influenced by both short-term and long-term preferences. Another key challenge in SRS lies in the data sparsity problem. The available dataset contains only partial observation of reality, not to mention the interaction data in a period of time, due to practical reasons such as privacy concerns or interaction in different platforms. My research problem is to capture users' long-term preferences and address the data sparsity problem in SRS. I propose to employ transfer learning and curiosity learning. First, I propose to transfer the knowledge of users' long-term preferences to the SRS by initializing user and item embeddings with the whole historical interaction information. I propose a new initialization scheme called Laplacian Eigenmaps with Popularity-based Regularization for Isolated Data (Leporid). Leporid endows the embeddings with information regarding multi-scale neighborhood structures on the data manifold and performs adaptive regularization to compensate for high embedding variance on the tail of the data distribution. Exploiting matrix sparsity, Leporid embeddings can be computed efficiently. I evaluate Leporid in a wide range of recommendation models, including both sequential and nonsequential ones, where Leporid improves the model accuracy with large margins. To maximize the effects of the initialization, I propose a sequential recommendation model, named Dual-Loss Residual Recommendation (DLR2), which, when initialized with Leporid, substantially outperforms both traditional and state-of-the-art neural recommendation systems. Second, transferring users' interaction information across different domains is one possible solution to solve the data sparsity problem because more interactive data are available and there usually exist sequential dependencies between users' interactive items in different domains. I propose a novel deep cross-domain sequential recommendation model, called Cycle Generation Networks (CGN). Specifically, CGN employs two generators to construct the dual-direction personalized itemset mapping between a user's behaviors in two different domains over time. The generators are learned by optimizing the distance between the generated itemset and the real interacted itemset, as well as the cycle-consistent loss which is defined according to the dual-direction generation procedure. Extensive experiments on real datasets demonstrate the effectiveness of the proposed model, comparing with existing single-domain and cross-domain recommendation methods. Third, I propose to employ curiosity based conversations in SRS when faced with limited user-item interaction data. Inadequate data usually result in uncertain predictions. When the SRS is uncertain about modeling user preferences, it may be advantageous to directly ask users for their preferences over item attributes in the form of conversations. I propose a simple and efficient conversational sequential recommendation method, MInimalist Non-reinforced Interactive COnversational Recommender Network (Minicorn). Minicorn models the epistemic uncertainty of the estimated user preference and queries the user for the attribute with the highest uncertainty. The system employs a simple network architecture and makes the query-vs-recommendation decision using a single uncertainty-based rule. Somewhat surprisingly, this minimalist approach outperforms state-of-the-art reinforcement learning methods on three real-world datasets by large margins.