Export
Xiangwei Lv
· 2 previous affiliations
- Zhejiang UniversityCurrent
- Current
- PhD student · 2024–2026
- MS student · 2021–2024
- Tongji University2017–2021
- Undergrad student · 2017–2021
Research Overview
Xiangwei Lv focuses on graph neural networks, out-of-distribution detection, and representation learning for educational data mining. Recent work addresses the challenges of domain adaptation and semantic shifts in text-attributed graphs, utilizing large language models to improve model robustness. Current research emphasizes debiasing cognitive representations and improving knowledge transfer across graph domains.
Research Trajectory
Xiangwei Lv’s research trajectory is characterized by a focus on improving the robustness, transferability, and reliability of machine learning models, particularly when dealing with graph-structured data and human-centric behavioral data. Operating primarily as a first author, Lv has led several recent investigations into how models handle data that deviates from their initial training distributions, whether through domain shifts, out-of-distribution anomalies, or noisy behavioral signals.
A major focal point of Lv’s work is the application and refinement of Graph Neural Networks (GNNs). While GNNs have become the standard for learning from relational data, they frequently struggle with generalization and domain gaps. Lv has addressed these limitations by exploring how models can better handle text-attributed graphs (TAGs) when deployed in unpredictable environments. In the 2025 paper Out-of-Distribution Detection via LLM-Guided Outlier Generation for Text-attributed Graph, Lv and collaborators tackled the vulnerability of TAG prediction models to out-of-distribution (OOD) samples. Because high-quality OOD data is notoriously scarce and difficult to curate, training robust detectors is a significant challenge. As the first author, Lv proposed a novel methodology that leverages Large Language Models (LLMs) to systematically generate outliers, thereby regularizing the detection model without relying on exhaustive manual datasets. This work contributes directly to the broader area of Distinguishing Semantically Similar Outliers from Known Classes, providing a mechanism to help models recognize when they are encountering data that falls outside their reliable operational parameters.
In parallel with OOD detection, Lv has investigated the challenge of transferring learned graph representations across different domains. Standard GNNs trained on a specific source domain often experience a sharp degradation in performance when applied to a target domain due to inherent structural and feature-based domain gaps. In Grasp the Key Takeaways from Source Domain for Few Shot Graph Domain Adaptation (2025), Lv led research into Few-Shot Graph Domain Adaptation (FSGDA) for node classification tasks. The research addresses the pitfalls of naively co-training models on fully labeled source graphs and sparsely annotated target graphs, a practice that often leads to suboptimal knowledge transfer. By focusing on extracting and adapting only the most relevant "key takeaways" from the source domain, this work improves the efficiency and accuracy of domain adaptation. This focus on optimizing GNN performance and transferability aligns with ongoing challenges in graph learning, including themes such as the research direction Distilling Graph Neural Networks into Graph Free Models, which similarly seeks to overcome the structural and computational bottlenecks of traditional message-passing architectures.
Beyond graph-specific architectures, Lv has applied advanced representation learning techniques to the domain of intelligent education, specifically focusing on Knowledge Tracing (KT). Knowledge tracing aims to model a student's evolving knowledge state over time to predict future performance. However, traditional KT models often operate on the flawed assumption that a student's recorded interactions perfectly reflect their underlying cognition. In the 2025 article Debiased Cognition Representation Learning for Knowledge Tracing, published in the ACM Transactions on Information Systems, Lv served as first author on a project designed to correct this assumption. The research highlights how abnormal online learning behaviors—such as guessing or plagiarism—introduce severe biases into training data, corrupting the model's assessment of the student. To resolve this, Lv introduced a debiased learning framework that disentangles genuine cognitive mastery from noisy behavioral artifacts, resulting in a higher-quality representation of student knowledge.
Current Directions
The most recent papers in the evidence indicate that Lv’s current research is heavily invested in the reliability of representation learning under conditions of uncertainty, bias, and data scarcity. A primary ongoing direction is the mitigation of distribution shifts in complex data structures. This involves developing more sophisticated methods for out-of-distribution detection in text-attributed graphs and refining few-shot adaptation techniques so that models can generalize to new domains with minimal manual annotation.
Additionally, Lv's recent work points to a sustained interest in algorithmic debiasing, particularly in human-centric applications like educational data mining. By identifying and filtering out misleading behavioral signals, this line of research seeks to build more accurate, equitable, and robust predictive models. Across both graph learning and knowledge tracing, the unifying theme of Lv's current work is the drive to make machine learning systems more resilient to the noisy, shifted, and imperfect data encountered in real-world deployments.
Profile updated
| Title | Venue | Year |
|---|---|---|
| Debiased Cognition Representation Learning for Knowledge Tracing | ACM Trans. Information Systems | 2025 |
| Grasp the Key Takeaways from Source Domain for Few Shot Graph Domain Adaptation | WWW | 2025 |
| Out-of-Distribution Detection via LLM-Guided Outlier Generation for Text-attributed Graph | ACL | 2025 |
Loading new papers…
Loading research neighbors…
See something wrong? Report a mistake