| Author: | Chen, Xingming |
| Title: | Causal representation learning for sequential user behaviors : from generic to news recommendation |
| Advisors: | Li, Qing (COMP) |
| Degree: | Ph.D. |
| Year: | 2026 |
| Department: | Department of Computing |
| Pages: | xii, 119 pages : color illustrations |
| Language: | English |
| Abstract: | The rapid expansion of digital platforms has rendered personalized recommendation systems essential for connecting users with pertinent information in increasingly vast and dynamic content environments. Among these systems, sequential recommendation (SR)—which models users' evolving interests based on their behavioral histories—serves as a cornerstone for a wide range of applications, including e-commerce and news delivery. Despite considerable advancements, the majority of existing methodologies are grounded in statistical associations, often overlooking the underlying causal mechanisms that govern user behaviors. This limitation frequently results in spurious correlations and various forms of bias, thereby undermining the robustness and interpretability of recommendation outcomes. This thesis systematically investigates causal representation learning for sequential user behaviors, with particular emphasis on bridging generic sequential recommendation and the distinct challenges inherent to news recommendation (NR). In this context, three fundamental research challenges are identified and addressed through a unified research agenda. First, most existing causality-based recommendation approaches rely heavily on explicit external signals, such as item popularity or display position, to deconfound user-item correlations. However, such features are frequently unavailable or insufficient in real-world scenarios. This raises the critical problem of how to uncover hidden confounders and spurious factors that induce bias in user behavior modeling when explicit signals are absent. To address this, a causality-driven sequential user modeling approach is introduced, which constructs causal graphs to identify and isolate unobservable confounding factors within user interaction histories. This framework effectively distinguishes genuine user interests from spurious correlations, thereby mitigating the impact of implicit biases and shortcut pathways in generic sequential recommendation tasks. The second challenge stems from the inadequacy of relying on predefined causal graphs in complex and dynamic recommendation environments. As user behaviors and item spaces grow increasingly diverse, static causal structures fail to capture the dynamic generative processes underlying user decision-making. To overcome this limitation, the focus shifts to news recommendation—a domain characterized by rapidly changing topics and low content similarity between consecutive items. To meet this challenge, a Causal Behavior Pattern Inference (CBPI) framework is developed, which learns latent causal graphs directly from user interaction data. This approach enables the model to flexibly adapt to the evolving landscape of user interests and news content. The third challenge lies in the necessity of capturing multi-level causality in user behavior, which includes both the dependencies among observed interactions and the generative influence of latent user interests on these behaviors. Existing causal discovery (CD) methods are typically restricted to a single variable layer, rendering them inadequate for modeling the hierarchical and intertwined nature of real-world user decision processes. To this end, a hierarchical causal-inspired structure modeling method for news recommendation is proposed, which jointly uncovers causal relationships between latent interests and observable behaviors, resulting in more interpretable and robust recommendations. By systematically addressing these three challenges, this thesis presents a unified causal representation learning framework that reconstructs user and item representations based on the underlying causal processes governing user behaviors. Extensive experiments on large-scale real-world datasets demonstrate that the proposed causality-aware approaches significantly enhance recommendation accuracy, diversity, and robustness, thereby establishing a new direction for the development of personalized recommendation systems that are both effective and explainable. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/14400