The Department of Data Science (DS) at CityUHK is delighted to announce that PhD student Maolin WANG has been awarded the Best Paper Award Runner-Up in the Applied Data Science Track at ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) 2025, one of the world's most prestigious conferences in knowledge discovery and data mining. The achievement was announced at the conference held in Toronto, Canada, from August 3-7, 2025.
Maolin WANG, a PhD candidate at CityUHK's DS Department, was recognized for his groundbreaking paper "Put Teacher in Student's Shoes: Cross-Distillation for Ultra-compact Model Compression Framework". The research was conducted under the supervision of Professor Xiangyu ZHAO from DS at CityUHK, in collaboration with researchers from AntGroup.
The award-winning research addresses critical challenges in deploying AI models on resource-constrained mobile devices, achieving a remarkable breakthrough by creating the smallest known BERT-based language model at just 1.91 MB. This ultra-compact size enables seamless deployment on mobile devices while maintaining robust performance, making it particularly valuable for applications demanding both efficiency and versatility.
At the heart of this achievement is the team's innovative framework, which introduces a novel cross-distillation method that fundamentally reimagines how large AI models transfer knowledge to compact versions. The cross-distillation approach positions teacher models to understand student perspectives, ensuring efficient knowledge transfer through parameter integration and mutual interplay between models. This is combined with a comprehensive compression pipeline that achieves unprecedented model size reduction while maintaining performance.
The research has already demonstrated significant real-world impact through deployment across multiple scenarios within the Alipay ecosystem. Since January 2024, the ultra-compact model has been serving 8.4 million daily active devices in live edge recommendation systems, achieving a 4.23% increase in click-through rates for coupon recommendations. Additionally, the framework has enabled substantial improvements in latency reduction and privacy-preserving capabilities for various applications serving millions of users.
"This achievement demonstrates how academic research can directly address real-world challenges in mobile computing," said Professor Xiangyu ZHAO. "By enabling sophisticated AI capabilities on edge devices while preserving privacy and ensuring real-time responsiveness, our work opens new possibilities for personalized, secure AI applications."
KDD is the premier international conference bringing together researchers and practitioners from data science, machine learning, and AI. The Applied Data Science Track specifically focuses on research with real-world impact, making this recognition particularly meaningful for the team's industry-academic collaboration. This achievement reinforces CityUHK's position as a leading institution in AI and data science research, demonstrating the department's commitment to developing innovative solutions that bridge theoretical advances with practical applications. It also highlights the success of collaborative research between academia and industry partners in addressing complex technological challenges that benefit society at large.