Multi-task learning models of time-series sensing data for personalized context recognition -

Abstract

Time-series data is increasingly collected from a variety of real-world sources such as stock prices, power consumption, and vital signs from wearable sensors. With advances in machine learning, there has been a growing demand for predictions from time-series data. Accurate predictions here often entail the development of personalized models that capture the unique characteristics of an entity and, importantly, the need for large training datasets with informative content for the machine learning model. From a practical perspective, however, datasets that are unique to a specific entity are often not large enough particularly (i) when historical personalized data is naturally scarce or sparse, as in medical diagnostics and treatments of different diseases, or (ii) since active collection of labeled data that is unique to an entity is costly in terms of time and effort. To overcome the availability of limited data, we herein propose two multitask machine learning models that aim to increase the prediction accuracy of personalized modeling. To derive the desired designs, we propose the evaluation of three components to characterize the performance of multitask learning models for time-series, which include: time-series features that capture temporal dynamics, similarity measures that learn commonality and uniqueness elements across entities, and regularization factors that enhance model generalizability. The first model we propose is a multitask feature-based learning with time-series (MTFL-TS) model that uses Conditional Random Fields (CRF) with dynamic feature model representations of the time-series data, group-level similarity based on Principal Component Analysis (PCA), and various measures of regularization. The second design is a hierarchical multitask deep learning with time-series (MTDL-TS) model using a Convolutional Neural Network (CNN) to extract optimal features and commonality across entities and a Gated Recurrent Unit (GRU) to extract shared and unique characteristics of the temporal dynamics of eac