Donghong Cai (蔡东鸿)

About Me

Hello! My name is Donghong Cai, currently a PhD student at Washington University in St. Louis majoring in Computer Science. I earned my master's degree at Zhejiang University, with the privilege of being advised by Prof. Yang Yang, and my bachelor's degree from Huazhong University of Science and Technology. I also worked as a full-time machine learning engineer at Alibaba for over a year.

My research interests primarily include Machine Learning and Data Mining, with a particular focus on time series and graph, and it's applications in healthcare, social media, etc. During my tenure with Alibaba, I also developed keen interests in LLMs, and Foundation Models (in time series and graph).

Research Projects

Epileptic Wave Detection Using Hierarchical Graph Diffusion Learning

• Formulated and studied the epileptic wave detection problem for SEEG data using an automatic end-to-end data-driven method.

• Proposed BrainNet to simultaneously learn the dynamic diffusion graphs and model the brain wave diffusion patterns thereon in a hierarchical fashion to achieve accurate epileptic wave detection under conditions of imbalanced labels and severe noise.

• Conducted comprehensive experiments on a large-scale real-world SEEG dataset across multiple patients. The experimental results validated the effectiveness of BrainNet on epileptic wave detection and its superiority in capturing the diffusion process. In the channel-level epileptic wave detection task, BrainNet outperforms all baselines on F2-score with an increase of 36.66%.

Pre-training Framework for Brain Signals

• Designed a generalized self-supervised learning framework MBrain consisting of three well-designed tasks, which can be applied to pre-train both EEG and SEEG brain signals.

• Proposed multi-channel CPC (Contrastive Predictive Coding) and theoretically proved that optimizing the goal of multi-channel CPC can lead to a better predictive representation. Based on the multi-channel CPC, three selfsupervised learning tasks were designed to explicitly capture the spatial and temporal correlations of brain signals to learn informative representations for downstream tasks.

• Validated the superior effectiveness and clinical value of the proposed framework through extensive experiments of seizure detection on large-scale real-world EEG and SEEG datasets. In subject independent seizure detection experiment which meet practical clinical needs, MBrain outperforms all baselines on F2-score with an increase of 9.23% and 27.83% in EEG and SEEG datasets.

Multimodal Learning with Graph Alignment

• Introduced the graph modality into the realm of multi-modal fusion and proposed a graph alignment task to effectively combine multiple modalities.

• Proposed MMGA (Multi-Modal learning with Graph Alignment), an innovative pre-training framework designed to unify information from graph (social network), image, and text modalities on social media platforms to enhance user representation learning.

• Constructed the first multi-modal social media dataset containing image, text, and graph modalities, including over 2 million posts, a million-scale graph, and labeled user/post tags.

Industrial Projects (Selected)

Multiple Treatments Uplift Modeling for Voucher Recommendation

• Researched and designed the store promotional voucher recommendation algorithm on the LAZADA e-commerce platform from the perspective of individual treatment effect estimation in the causal inference domain.

• Defined a new multiple treatments uplift modeling problem in this scenario and proposed a deep learning-based model M-DESCN to address it. Conducted experiments on the dataset sampled and generated from the stores’ historical voucher data, demonstrating the effectiveness of the method from the experimental results on the offline verification set.

• Summarized the deficiencies in the current generated dataset and methodology, and contemplated potential solutions for further optimization in the future.

Same-store Identification Framework

• Designed an End-to-end cross-platform framework for identifying the same e-commerce sellers based on Innovative image data augmentation techniques focusing on semantic feature learning for Logo data. Our accuracy is up to 90% and resulting in a cost savings of $73,000 for one quarter.

E-commerce Merchant Platform‘s Chatbot Construction

• Involved in implementing an e-commerce merchant platform‘s question-answering bot.

• Employed QA-pair matching and the Retrieval-Augmented Generation (RAG) approach to provide domain knowledge to Language Models (LLMs) beyond their trained parameters.

• Enhanced our understanding of queries, extract corresponding values, and search for relevant information in merchant and product databases.

About Me

Publications

Research Projects

Industrial Projects (Selected)

Education

Professional Experience