Classical machine learning and deep neural network ensemble model for GPS-based activity recognition

Abstract

Our KDDI Research team proposes an ensemble model for the Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge using GPS data. During preprocessing, we corrected the GPS dataset that contains errors and missing values using a Kalman smoother. Since this smoother can be processed offline, it can be used to correct backforward using the recorded future locations in the dataset. Further, by using features that have similar distributions in the train subject's data and the other subjects' datasets, we could achieve robust feature selection across multiple subjects. The first stage of our ensemble model employs classical machine learning and deep neural network approaches independently, specifically, LightGBM and LSTM, respectively. The second stage calculates the weighted average of the outputs of both approaches. Our results show the improved accuracy contributed by our ensemble, suggesting that it effectively makes use of both statistical and nonstatistical features given the suitable base models. We confirmed that the use of Kalman-smoother, selection of features with similar distributions across subjects and ensemble modeling contributed to improving the accuracy of both train and validation datasets.