CityUHK SDSC Students Shine at KDD Cup 2024

CityUHK SDSC Students Shine at KDD Cup 2024

The School of Data Science at CityUHK is proud to announce the outstanding achievement of its students in the Amazon KDD Cup 2024: Multi-Task Online Shopping Challenge for LLMs. A team of five talented students secured an impressive second place overall (2/508), with top three finishes across all individual tracks of the competition.

News on KDD Cup 2024

 

The team comprises three Ph.D. candidates from the School of Data Science: JIA Pengyue, GAO Jingtong, and LI Xiaopeng, as well as two undergraduate students: JIN Yiyao from Data Science and WANG Zixuan from Computer Science. Their participation was under the guidance of Professor ZHAO Xiangyu from the School of Data Science, who is the director of the research group Applied Machine Learning Lab (AML Lab).

KDD Cup is a renowned competition that brings together excellent minds in data science to tackle real-world challenges using advanced machine learning techniques. This competition, Amazon KDD Cup 2024, brought together 508 teams, 1499 entrants from enterprises, universities and research institutes around the world. This year's challenge focused on multi-task learning for online shopping, providing a platform for participants to showcase their ability to apply large language models (LLMs) in diverse, complex scenarios.

This year’s competition is challenging due to the heterogeneity of the shopping scenarios, the complexity of tasks, the constrains of time and memory, and the absence of training data. To tackle these challenges, the team designed a comprehensive pipeline:
1.     Model Selection: To adapt to the competition's constraint of no training data, the team selected
         pre-trained large language models as the base model, tested and selected the best among them on
         development datasets.
2.     Quantization: The team uses quantization techniques, retaining as much of the large language model’s
         capabilities as possible under the constraints of time and memory.
3.     Prompt Design: To enhance the model's response quality across various scenarios and tasks, the team             design and refine prompts separately for different tasks and scenarios (e.g., different instructions and
         examples for different tasks and scenarios).

In the highly competitive Amazon KDD Cup 2024, the team achieved remarkable rankings:

  • Track 1-Understanding Shopping Concepts: 2nd Place
  • Track 2-Shopping Knowledge Reasoning: 3rd Place
  • Track 3-User Behavior Alignment: 3rd Place
  • Track 4-Multi-Lingual Abilities: 3rd Place
  • Track 5-All-Around: 2nd Place

This remarkable achievement by the CityUHK team demonstrates the University’s commitment to fostering talent and driving innovation in the field of data science. It also reinforces CityUHK’s position as a leader in nurturing future leaders in technology and research.