Congrats, 本科小朋友研究工作被BIBM2025接收

泽宇为第一作者，留学生Javeed和本科组员刘枭等参与的研究工作《“大脑连接结构阐释和知识蒸馏的图级异常检测”/Graph-level Anomaly Detection of Brain Connectivity with Structure Interpretation and Knowledge Distillation》被BIBM接收。感谢何老师联合指导。

论文摘要： Graph-level anomaly detection is a critical yet underexplored task, especially in the neuro-imaging domain the early identification of abnormal brain patterns is vital. In this paper, we propose KDGAE, a lightweight and generalizable graph autoencoder framework based on knowledge distillation. The architecture consists of a powerful teacher encoder and a simplified student encoder, trained to reconstruct input graphs and mimic latent representations. Anomaly scores are derived by jointly assessing reconstruction errors and embedding discrepancies. To enhance interpretability, we incorporate post hoc explanation techniques, including feature masking and GNNExplainer, enabling node-level attribution of global anomalies. We evaluate KDGAE on the UB-GOLD benchmark suite (15real-world graph datasets) and the Autism Brain Imaging DataExchange (ABIDE) dataset for autism detection. Our method demonstrates competitive performance against state-of-the-art baselines, particularly under limited supervision. Notably, the explanation results align with known neurological markers of Autism Spectrum Disorder (ASD), underscoring the biological plausibility of KDGAE.

图级异常检测是一项关键但尚未得到充分探索的任务，尤其在神经影像领域，早期识别异常脑模式至关重要。本文提出KDGAE，一种基于知识蒸馏的轻量级可泛化图自编码器框架。该架构包含一个强大的教师编码器和一个简化的学生编码器，通过训练实现输入图的重构与潜在表征的模仿。通过联合评估重构误差与嵌入差异得出异常评分。为提升可解释性，我们整合了特征掩码与GNNExplainer等事后解释技术，实现全局异常的节点级归因分析。我们在UB-GOLD基准测试集（含15个真实世界图数据集）和自闭症脑成像数据交换（ABIDE）数据集上评估KDGAE。实验表明，该方法在有限监督条件下尤其优于当前最优基线模型。值得关注的是，解释结果与自闭症谱系障碍（ASD）的已知神经标记物高度吻合，印证了KDGAE的生物学合理性。