Abstract
dc:description.abstractUser-generated textual data such as reviews, comments, posts, and item descriptions are increasingly prevalent across online platforms. These texts contain rich semantic evidence about user preferences and intentions that often goes beyond what can be inferred from ratings or clicks alone. However, transferring models across domains remains challenging: language, style, and semantics shift across platforms and product categories, supervision is often sparse, and cold-start users may have no interaction history in a new domain. In addition, as cross-domain models become more complex, it becomes difficult to audit their behavior and to ensure that their explanations are faithful to the underlying prediction mechanism. This thesis investigates cross-domain learning through a text-centric lens, focusing on two core problem families: multi-domain text classification under limited supervision and cold-start cross-domain recommendation. We develop four complementary frameworks that form a coherent progression from robust transfer to structured preference modeling and finally to verifiable explanations. First, we address multi-domain text classification under limited supervision by coupling supervised contrastive learning with domain-adversarial training in DaCon, our framework for learning label-discriminative yet domain-agnostic representations that generalize reliably to unseen or weakly labeled domains. Second, we tackle cold-start cross-domain recommendation by synthesizing auxiliary target-domain review signals from like-minded users, aligning user-item representations via supervised contrastive objectives, and enforcing domain invariance through adversarial adaptation in OmniMatch, our review-based transfer framework. Third, we improve transfer robustness when overlapping users are scarce by replacing brittle continuous embedding alignment with hierarchical discrete preference abstractions in RQ-Rec, which uses large language models to rewrite noisy review histories into concise, interest-focused summaries and then applies residual vector quantization to encode user preferences into multi-level codewords that can be mapped across domains. Finally, we address auditability by distilling reviews into evidence-cited facet cards and predicting in an evidence-grounded concept space with a linear transfer and an additive scoring function in EviSnap, yielding exact per-concept contributions and faithful, directly testable explanations. Collectively, this thesis demonstrates that transforming heterogeneous text into transferable, structured, and evidence-grounded representations can yield cross-domain systems that are both accurate and transparent.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy (Ph.D.)
- Level thesis:degree_level
- Doctoral
- Discipline thesis:degree_discipline
- Computer Science
- Grantor dc:publisher
- Carleton University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Dai, Yingjun
Subjects
dc:subject × 2Rights
dc:rights- Statement dc:rights
-
- Copyright © 2026 the author(s). Theses may be used for non-commercial research, educational, or related academic purposes only. Such uses include personal study, distribution to students, research and scholarship. Theses may only be shared by linking to the Carleton University Institutional Repository and no part may be copied without proper attribution to the author; no part may be used for commercial purposes directly or indirectly via a for-profit platform; no adaptation or derivative works are permitted without consent from the copyright owner.
- Language dc:language.iso
- en