| Author: | Fan, Lu |
| Title: | User intent analysis and behavior modeling for conversation and recommendation |
| Advisors: | Wu, Xiao-ming (DSAI) |
| Degree: | Ph.D. |
| Year: | 2026 |
| Department: | Department of Data Science and Artificial Intelligence |
| Pages: | xii, 143 pages : color illustrations |
| Language: | English |
| Abstract: | Understanding and modeling user intent and behavior are foundational challenges in the development of intelligent conversational AI and recommender systems. As dialogue systems and personalized recommendation platforms become increasingly prevalent, the ability to accurately identify user intentions and effectively capture user behavior patterns is critical for delivering adaptive, user-centric services. However, these systems face significant hurdles: conversational AI must contend with the rapid emergence of new user intents, meanwhile, recommender systems must model complex, evolving user preferences from behavioral data. Motivated by these challenges, this thesis explores two complementary research directions: (a) user intent analysis in conversational systems, and (b) user behavior modeling for recommendation. In the first direction, we advance the state of intent detection, with a particular focus on zero-shot and unknown intent classification. We propose novel methods to reconstruct capsule networks for zero-shot intent classification, introducing dimensional attention mechanisms to address polysemy and leveraging latent information to improve generalization to unseen intents. Additionally, we develop a semantic-enhanced Gaussian mixture model for unknown intent detection, integrating it with zero-shot intent classifiers to further boost performance. Recognizing the limitations of existing approaches in new intent discovery, we present LANID, a framework that harnesses the semantic power of large language models to guide lightweight encoders, enabling scalable and effective identification of novel intents in task-oriented dialogue systems. In the second research direction, we address the challenge of modeling user behavior for personalized product search and sequential recommendation. We propose graph-based approaches that capture both local and global user behavior patterns, utilizing user successive behavior graphs and efficient graph convolution techniques to enrich product representations and uncover collaborative signals. Furthermore, we introduce a neighborhood-based hard negative mining strategy for sequential recommendation, exploiting structural information in user-item interaction graphs to enhance negative sampling and improve model training. Extensive experiments across multiple real-world datasets and benchmarks demonstrate the effectiveness and generalizability of our proposed methods, consistently outperforming strong baselines in both conversational AI and recommender system tasks. By tackling the fundamental problems of intent recognition and user behavior modeling, this thesis contributes novel solutions that advance the adaptability, intelligence, and user-centricity of next-generation conversational and recommendation systems. |
| 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/200/14647