A systematic approach to multi-task learning from time-series data

Abstract There have been significant advances in machine learning due to the profusion in data collection and computing resources. However, the need for large annotated datasets to train machine learning models remains a problematic constraint. To address the limitation of annotated data for personalized prediction, we propose a framework to enrich annotated time-series (TS) sensing data by way of transfer learning with new multi-task learning (MTL) models. Compared to previous MTL approaches for TS, this work introduces three contributions. First, we propose a systematic method to examine the efficiency of MTL approaches by exploring options for three key characteristics of MTL with TS: the choice of features that efficiently capture temporally dynamic information, the similarity measure that effectively models the commonality and uniqueness across tasks being learned, and the choice of regularization for achieving the best tradeoff between a model’s generalizability and accuracy. Second, we present an MTL deep learning model that is shown to achieve state-of-the-art performance for personalized human activity recognition from time-series. Experimental results on three benchmark activity recognition datasets and one activity recognition in-the-wild dataset show that the proposed framework provides performance gains over prior work while presenting a unified approach for designing MTL solutions for personalized time-series classification problems.