Detailed instructions for how to use our pipeline are provided in ‣. This document focuses only on the approach and results.
The full runtime of submit_onlinetrain.ipynb on the platform is 2h 11min (training + inference), and the AUC score on the Public Leaderboard is 88.62%.
Chinese version Solution:
Competition Link:
Code Repository:
First and foremost, we would like to express our heartfelt gratitude to CrunchDAO and ADIA Lab for organizing this competition. This is the second time we participated in ADIA’s competition, our sincere thanks go to all the staff involved for their hard work in creating such an exciting and challenging event.
This solution was a collaborative effort by Mutian Hong from ShanghaiTech University and Guoqin Gu from Xiamen University. We are both undergraduate senior students; Mutian will pursue a direct PhD program and Guoqin will pursue a Master’s degree.
Contract Us:
The task is to determine whether there is a structural break in a time series before and after a given timestamp. The time series is split into two segments by that timestamp, left (period = 0) and right (period = 1), and we aim to extract as many informative features as possible.
At first, we tried a CrossEncoder approach that treats the left and right segments as two “sentences.” Inspired by PatchTST’s patching idea, we converted the continuous time-series signal into tokens, turning the task into an NLP-style sentence similarity task. Unfortunately, the AUC of this approach was close to random.
Next, we tried UTime, a model for sleep signal segmentation, but the results were still unsatisfactory.