Export
Yangyang Wu
· 2 previous affiliations
- Zhejiang UniversitySince 2019
- Assistant Professor · Since 2023
- PhD student · 2019–2023
- Zhejiang University of Technology2016–2019
- MS student · 2016–2019
Research Overview
Yangyang Wu focuses on data imputation, the adversarial robustness of graph neural networks, and multimodal machine learning. His early work explored network security and generative models for handling missing data in complex datasets. More recently, his research addresses parameter-efficient adaptation and robust inference for multimodal systems facing incomplete or unreliable information.
Research Trajectory
Early in his career, Wu contributed to distributed systems research, working on A New Self-Adapting Knowledge Fusion System (2007), which utilized genetic algorithms and ontology technology to optimize knowledge fusion with a self-adapting feedback mechanism.
By 2018, his focus shifted toward the security and vulnerabilities of graph-based machine learning models. He investigated the privacy risks of network embeddings in Fast Gradient Attack on Network Embedding, proposing a framework to generate adversarial networks using gradient information from Graph Convolutional Networks (GCNs). This investigation into model fragility led to Can Adversarial Network Attack be Defended? (2019), a study exploring how targeted structural perturbations can deceive Graph Neural Networks (GNNs) and how robust defenses might be constructed. This work contributed to the research direction of Scalable Adversarial Robustness for Graph Neural Networks, as well as broader areas such as The Fundamental Tradeoff Between Robustness and Performance and Regularizing Neural Networks via Label Smoothing. Continuing this trajectory, he analyzed defense strategies in Smoothing Adversarial Training for GNN (2021), which applied adversarial training techniques to graph mining to mitigate the impact of crafted perturbations.
Concurrently, Wu began an extensive line of research addressing the pervasive challenge of missing data in complex datasets. In 2021, he designed a semi-supervised generative adversarial network for handling incomplete multivariate time series in Generative Semi-supervised Learning for Multivariate Time Series Imputation. He also explored alternative imputation methodologies, such as the use of influence functions in Efficient and Effective Data Imputation with Influence Functions.
To make these advanced techniques more accessible and scalable, Wu introduced An Interactive Data Imputation System (2022), a platform integrating multiple state-of-the-art algorithms and a novel variational generative network to allow users to customize imputation pipelines. He subsequently expanded on generative approaches with Jointly Imputing Multi-View Data with Optimal Transport (2023), proposing the Git model to handle missing features across multi-view data using optimal transport theory. To address the computational bottlenecks of large-scale incomplete data, he presented the SEMI toolbox in An Efficient Generative Data Imputation Toolbox with Adversarial Learning. His work on database efficiency also extended to query optimization, as seen in his research on Efficient and Effective Cardinality Estimation for Skyline Family (2023), which introduced an end-to-end model for predicting query result sizes over massive databases.
Consolidating his contributions to generative imputation, Wu detailed the SCIS system in Differentiable and Scalable Generative Adversarial Models for Data Imputation (2024). This work accelerated GAN training by introducing a masking Sinkhorn divergence, significantly reducing computational costs while maintaining accuracy. To provide a comprehensive view of the field, he also contributed to An Experimental Survey of Missing Data Imputation Algorithms (Extended Abstract), systematically comparing 19 algorithms across numerous benchmark datasets.
In 2025, Wu's research expanded into multimodal machine learning and the integration of large language models (LLMs) for data processing. He contributed to A Wander Through the Multimodal Landscape: Efficient Transfer Learning via Low-rank Sequence Multimodal Adapter, which introduced a parameter-efficient adapter method for fine-grained, token-level interactions across multiple modalities. He also explored the reasoning capabilities of LLMs for data cleaning, participating in the design of the ZeroED framework in ZeroED: Hybrid Zero-Shot Error Detection Through Large Language Model Reasoning.
Returning to the challenge of incomplete information, Wu applied his expertise in missing data to both tabular and multimodal domains. He proposed MISS: An Incomplete Tabular Data Representation System with Missing Mechanism Learning (2025), a system that incorporates missingness propensity scores to enable Transformer-based models to handle incomplete tabular data effectively. In the multimodal space, he addressed the dynamic loss of sensor streams or data channels in Efficient Prompting for Continual Adaptation to Missing Modalities, formulating the issue as a continual learning task. This work advances the research area of Robust Multimodal Inference with Arbitrary Modality Subsets. Finally, his research group addressed test-time adaptation in Bridging the Gap for Test-Time Multimodal Sentiment Analysis, focusing on maintaining model reliability during deployment.
Current Directions
Recent work investigates the intersection of missing data mechanisms and multimodal machine learning. The most recent papers in the evidence address the challenges of deploying models in environments with incomplete information, such as handling dynamic missing modalities in continual learning settings and representing incomplete tabular data for Transformer architectures. Additionally, current research explores parameter-efficient transfer learning for multimodal foundation models and the use of large language models for zero-shot error detection in data preparation pipelines.
Profile updated
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake