Triple-O for SHL Recognition Challenge: An Ensemble Framework for Multi-class Imbalance and Training-Testing Distribution Inconsistency by OvO Binarization with Confidence Weight of One-class Classification

Abstract

SHL Recognition Challenge provides phone sensor data for recognizing eight modes of locomotion and transportation(activities). Most motion sensor data is user-dependent, so the model's generalization ability requires various data at the user level. The SHL dataset is collected only by three users, and category imbalance and distribution gap between training data and testing data can be significant obstacles for this challenge. In comparison to exploring the large and deep model structure, some model-free tricks play a more critical role in the previous challenges, such as transfer learning, re-sampling and anti-overfit trick. SHL challenge also notices that problem and provides user-independent sensor data in 2021. This paper analyses the data and finds out that category imbalance and distribution inconsistency are still the obstacles. This paper focuses on evaluating different methods to improve the general predicting ability at the setting of category imbalance and training-Testing distribution inconsistency. Besides, this paper puts forward a new ensemble framework called triple-O, using OvO binarization and one-class classification. The results show that the OvO binarization ensemble gets better results on the hard-To-distinguish class than re-sampling and re-weighting. One-class classification can be an anomaly detection to re-weight the meta learners to tackle the distribution gap. Triple-O can be plug-And-play and pave the way for exploring complex model structures. This paper introduces the solutions from the team GoodGoodDriveDayDayUp.