Wen Shen (沈雯)

Wen Shen (沈雯)

Associate Professor, Tongji University

Publications

TME-PSR: Time-aware, Multi-interest, and Explanation Personalization for Sequential Recommendation

Published in arXiv preprint, 2026

Abstract. This paper proposes a sequential recommendation model that integrates time-aware personalization, multi-interest personalization, and explanation personalization to capture temporal rhythm, fine-grained latent interests, and personalized semantic alignment.

Recommended citation: Wang, Q., Wen, L., Chen, J., Peng, K., Qin, R., Wei, Z., & Shen, W. TME-PSR: Time-aware, Multi-interest, and Explanation Personalization for Sequential Recommendation. arXiv preprint arXiv:2604.09439, 2026. https://arxiv.org/abs/2604.09439

Understanding and Defending VLM Jailbreaks via Jailbreak-Related Representation Shift

Published in arXiv preprint, 2026

Abstract. This paper studies how visual modality can induce jailbreak-related representation shifts in vision-language models and proposes a defense that removes the jailbreak-related shift at inference time.

Recommended citation: Wei, Z., Li, Q., Ruan, J., Qin, Z., Wen, L., Liu, D., & Shen, W. Understanding and Defending VLM Jailbreaks via Jailbreak-Related Representation Shift. arXiv preprint arXiv:2603.17372, 2026. https://arxiv.org/pdf/2603.17372

Interpreting Arithmetic Reasoning in Large Language Models using Game-Theoretic Interactions

Published in NeurIPS, 2025

Abstract. This paper interprets arithmetic reasoning in large language models using game-theoretic interactions, quantifying interaction patterns encoded during forward propagation to explain how LLMs solve arithmetic problems.

Recommended citation: Wen, L., Zheng, L., Li, H., Sun, L., Wei, Z., & Shen, W. Interpreting Arithmetic Reasoning in Large Language Models using Game-Theoretic Interactions. In NeurIPS 2025. https://openreview.net/pdf?id=tRvzEL64dY

A Unified Approach to Interpreting Self-supervised Pre-training Methods for 3D Point Clouds via Interactions

Published in CVPR, 2025

Abstract. This paper uses game-theoretic interactions as a unified approach to interpret self-supervised pre-training methods for 3D point clouds and identifies a shared mechanism behind their performance gains.

Recommended citation: Li, Q., Ruan, J., Wu, F., Chen, Y., Wei, Z., & Shen, W. A Unified Approach to Interpreting Self-supervised Pre-training Methods for 3D Point Clouds via Interactions. In CVPR 2025. https://openaccess.thecvf.com/content/CVPR2025/papers/Li_A_Unified_Approach_to_Interpreting_Self-supervised_Pre-training_Methods_for_3D_CVPR_2025_paper.pdf

Interpretable Rotation-Equivariant Quaternion Neural Networks for 3D Point Cloud Processing

Published in IEEE TPAMI, 2024

Abstract. This paper revises neural networks for 3D point cloud processing into rotation-equivariant quaternion neural networks, improving rotation robustness while preserving permutation invariance.

Recommended citation: Shen, W., Wei, Z., Ren, Q., Zhang, B., Huang, S., Fan, J., & Zhang, Q. Interpretable Rotation-Equivariant Quaternion Neural Networks for 3D Point Cloud Processing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. https://ieeexplore.ieee.org/abstract/document/10384563

Batch Normalization Is Blind to the First and Second Derivatives of the Loss

Published in AAAI, 2024

Abstract. This paper proves that batch normalization blocks the influence of the first- and second-order terms of the loss on earlier layers under a Taylor-series perspective.

Recommended citation: Zhou, Z., Shen, W., Chen, H., Tang, L., Chen, Y., & Zhang, Q. Batch Normalization Is Blind to the First and Second Derivatives of the Loss. In AAAI 2024. https://ojs.aaai.org/index.php/AAAI/article/download/29978/31715

Interpreting Representation Quality of DNNs for 3D Point Cloud Processing

Published in NeurIPS, 2021

Abstract. In this paper, we evaluate the quality of knowledge representations encoded in deep neural networks (DNNs) for 3D point cloud processing. We propose a method to disentangle the overall model vulnerability into the sensitivity to the rotation, the translation, the scale, and local 3D structures. Besides, we also propose metrics to evaluate the spatial smoothness of encoding 3D structures, and the representation complexity of the DNN. Based on such analysis, experiments expose representation problems with classic DNNs, and explain the utility of the adversarial training.

Recommended citation: Shen W., Ren Q., Liu D., Zhang Q.: Interpreting Representation Quality of DNNs for 3D Point Cloud Processing. In: Proceedings of the 35th Conference on Neural Information Processing Systems (NeurIPS). (2021) https://proceedings.neurips.cc/paper/2021/file/4a3e00961a08879c34f91ca0070ea2f5-Paper.pdf

Interpretable Compositional Convolutional Neural Networks

Published in IJCAI, 2021

Abstract. The reasonable definition of semantic interpretability presents the core challenge in explainable AI. This paper proposes a method to modify a traditional convolutional neural network (CNN) into an interpretable compositional CNN, in order to learn filters that encode meaningful visual patterns in intermediate convolutional layers. In a compositional CNN, each filter is supposed to consistently represent a specific compositional object part or image region with a clear meaning. The compositional CNN learns from image labels for classification without any annotations of parts or regions for supervision. Our method can be broadly applied to different types of CNNs. Experiments have demonstrated the effectiveness of our method.

Recommended citation: Shen W., Wei Z., Huang S., Zhang B., Fan J., Zhao P., Zhang Q.: Interpretable Compositional Convolutional Neural Networks. In: Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI). (2021) https://arxiv.org/abs/2107.04474

Verifiability and Predictability: Interpreting Utilities of Network Architectures for Point Cloud Processing

Published in CVPR, 2021

Abstract. In this paper, we diagnose deep neural networks for 3D point cloud processing to explore utilities of different intermediate-layer network architectures. We propose a number of hypotheses on the effects of specific intermediate-layer network architectures on the representation capacity of DNNs. In order to prove the hypotheses, we design five metrics to diagnose various types of DNNs from the following perspectives, information discarding, information concentration, rotation robustness, adversarial robustness, and neighborhood inconsistency. We conduct comparative studies based on such metrics to verify the hypotheses. We further use the verified hypotheses to revise intermediate-layer architectures of existing DNNs and improve their utilities. Experiments demonstrate the effectiveness of our method.

Recommended citation: Shen, W., Wei, Z., Huang, S., Zhang, B., Chen P., Zhao P., & Zhang, Q. Verifiability and Predictability: Interpreting Utilities of Network Architectures for Point Cloud Processing. In CVPR 2021. https://arxiv.org/abs/1911.09053v3

3D-Rotation-Equivariant Quaternion Neural Networks

Published in ECCV, 2020

Abstract. This paper proposes a set of rules to revise various neural networks for 3D point cloud processing to rotation-equivariant quaternion neural networks (REQNNs). We find that when a neural network uses quaternion features, the network feature naturally has the rotation-equivariance property. Rotation equivariance means that applying a specific rotation transformation to the input point cloud is equivalent to applying the same rotation transformation to all intermediate-layer quaternion features. Besides, the REQNN also ensures that the intermediate-layer features are invariant to the permutation of input points. Compared with the original neural network, the REQNN exhibits higher rotation robustness.

Recommended citation: Shen W., Zhang B., Huang S., Wei Z., Zhang Q.: 3D-Rotation-Equivariant Quaternion Neural Networks. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 531-547 (2020) https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123650528.pdf