Ensemble-based domain adaptation for transport mode recognition with mobile sensors

We present our submission (team S304) to the Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge 2019. The goal is to recognize 8 modes of transport and locomotion from 5 second frames of inertial sensor data of a smartphone carried in the hand, while most of the labelled data provided for classifier training consists of data from three other smartphone placements: hips, torso and bag. Only a small dataset from a smartphone carried in the hand was provided. Model training is complicated by the fact that the data distribution differs between the phone positions. To optimize classification performance for data from the Hand phone, we employ an ensemble of Multilayer Perceptrons, each trained with data from a different particular smartphone placement, including the small dataset of the Hand phone. We propose an iterative re-weighting scheme for combining the classifiers that takes their agreement with the specialized Hand classifier into account. The proposed method achieves 74% average per-class Recall, significantly improving the performance achieved when training with mixed data from all phone placements (59%) and training with data from the Hand phone only (66%). The ensemble-based method also outperforms domain adaptation by Feature Augmentation, which achieves 70% average Recall.