Towards trustworthy recommenders: building explainable and unbiased recommendation systems

• Main Author: Hu, Yidan
• Other Authors: Miao Chun Yan
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science
• Online Access: https://hdl.handle.net/10356/175790
• Institution: Nanyang Technological University

The explosively increasing online content, such as exposure on e-commerce platforms (e.g., Amazon and Taobao), makes it very difficult for users to choose suitable items or information from the vast volume of options available. To address this problem, recommendation systems have been widely used to assist users in making decisions by recommending a list of relevant items or information. Recommendation systems are powerful tools, but their lack of transparency in recommendation results and potential for unfair exposure to specific items, such as popular items, may harm users’ trust and satisfaction in the systems. As a result, there is a growing demand for trustworthy recommendation systems (TRS) that not only provide accurate recommendations but also adhere to key principles of trustworthiness. TRS aims to build a responsible and trustworthy recommendation system by considering five key aspects of trustworthiness: explainability, fairness, privacy, robustness, and controllability. In this thesis, we focus on two aspects of trustworthiness: explainability and fairness, which have been widely studied in TRS. Our goal is to develop an explainable and unbiased recommendation system that improves the transparency and fairness of recommendation systems, as unfair recommendations are caused by various biases. However, the construction of such a system still presents several challenges: 1) Existing explainable recommendation methods do not explicitly consider user preferences on different item aspects, which might lead to generating incoherent or unrelated explanations due to a lack of generative signals. 2) Existing long-tail solutions apply a "Closed-book" training strategy, which limits the model's ability to learn knowledge about tail items during training, resulting in poor performance on tail items. The "Closed-book" training strategy means that the model learns knowledge from training data by updating parameter weights, which are then used to recommend new items without extra information in the inference stage. However, due to the nature of learning objective functions, models often pay more attention to head items and learn little knowledge about tail items. 3) Conventional sequential recommendation models trained on long-tail data often produce biased recommendation results that provide more exposure to popular items (i.e., head items). One main reason is that rare items (i.e., tail items) acquire weak representation due to data sparsity. Some existing sequential recommendation methods mitigate the long-tail problem by enhancing representations of tail items, but these methods fail to explicitly enhance item representations due to ignoring the influence of context items. Consequently, these approaches encounter challenges in effectively and precisely enhancing the representation of items. To address the first issue, we propose an Aspect-guided Explanation generation with Syntax Graph (AESG) approach for explainable recommendations. AESG employs a review-based syntax graph to present a comprehensive perspective of user or item details. Additionally, an aspect-guided graph pooling operator is introduced to extract aspect-relevant information from the review-based syntax graphs to model the user's preferences on an item at the aspect level. Then, an aspect-guided explanation decoder is developed to generate aspects and aspect-relevant explanations based on the attention mechanism. The experimental results on three real-world datasets indicate that AESG outperforms state-of-the-art explanation generation methods in both single-aspect and multi-aspect explanation generation tasks and also achieves comparable or even better preference prediction accuracy than strong baseline methods. To address the second issue, we introduce a novel sequential recommendation framework, named MASR (i.e., Memory Bank Augmented Long-tail Sequential Recommendation). MASR is an "Open-book" model that combines novel types of memory banks and a retriever-copy network to alleviate the long-tail problem. During inference, the designed retriever-copy network retrieves related sequences from the training samples and copies useful information as a cue to improve the recommendation performance on tail items. Two designed memory banks provide reference samples to the retriever-copy network by memorizing the historical samples appearing in the training phase. We conduct extensive experiments on five real-world datasets to demonstrate the effectiveness of the proposed MASR model. The experimental results show that MASR consistently outperforms baseline methods in terms of recommendation performance on tail items. To tackle the final challenge, we propose a novel sequential recommendation framework, named DCAIR (i.e., Decoupling Context and Attribute-aware Information for Long-tail Recommendations). DCAIR takes a dual-pronged approach to enhance item representations, especially for those tail items. Different from existing representation enhancement solutions, we propose a disentanglement module that operates at a fine-grained level. This module separates the sequential features into context and attribute features to represent the items more precisely, especially for the tail ones. Besides, we leverage a pre-trained prompt model to learn multiple attributes, enabling the integration of attribute-specific knowledge into the decoupled features. To better represent tail items, we also implement a reconstruction encoder to transfer knowledge from head items to tail items using the decoupled features and pre-trained representations. We conduct extensive experiments on four real-world datasets to demonstrate the effectiveness of DCAIR. The experimental results indicate that DCAIR significantly outperforms six state-of-the-art baseline methods in terms of recommendation performance on tail items.