Professor Cheng Hong, Vice-Chairman (Graduate) of the Department of Systems Engineering and Engineering Management at The Chinese University of Hong Kong (CUHK), and her research team have won the prestigious Best Paper Award (Research Track) from the Association for Computing Machinery’s Special Interest Group on Knowledge Discovery and Data Mining (ACM SIGKDD) International Conference 2023.

Their paper “All in One: Multi-task Prompting for Graph Neural Networks” beat more than 1,400 other research submissions from around the globe to claim the award. This marks the first time in the history of SIGKDD that researchers from Hong Kong have been awarded this prestigious accolade.

Advancing Artificial Intelligence (AI) Pre-Training and Prompting Models

Currently, the multipurpose AIs that have been developed are mainly trained on large language models, which means generating human-like text responses based on training from enormous amounts of language data. The industry has met difficulties in training AI models with massive and diverse sources of data that form networked relationships rather than linear ones.

“All in One: Multi-task Prompting for Graph Neural Networks” presents a novel prompt and pre-training model for AI to learn from non-linear data such as diverse graphs, including complicated network graphs in different geographical locations, networks of antigens and antibodies in virology, connections on social networks, and others. The novel AI pre-training model has been proven to boost the accuracy rate of AI models by 1% to 8% compared to current training methods.

The paper’s first author Dr Sun Xiang-guo, currently a postdoctoral research fellow at CUHK under the supervision of Professor Cheng, was honoured to win the Best Paper Award, remarking, “The new pre-training model is expected to foster the development of versatile, large-scale, general AI that can understand graphs, process graph prompts and output graph responses, providing complex geographical and traffic analysis for public transport policy, accelerating vaccine and drug development with biological information databases, and facilitating social network big data analysis.”

SIGKDD’s judging panel affirmed the contribution of the paper, commenting, “There is a pressing need to explore more effective and efficient ways to exploit pre-trained graph models. This paper addresses a significant problem by unifying the format of graph prompts and language prompts with the prompt token, token structure and inserting pattern. The experimental evaluation demonstrates that the proposed algorithm improves on existing algorithms. This work lays the foundation for upcoming general graph AI models, which may have a more profound impact on applications such as brain science.”