Environment-agnostic Effective Learning for Domain Generalization on IoT Time Series Data

Massive time series data from different IoT devices often follow different distributions, leading to the failure of machine learning methods on newly generated data. Domain generalization methods are widely used to address such issues. The common idea of domain generalization is to learn the stable features across different environments. This involves dividing the training dataset into different environments based on distinct features for classification tasks, such as the rotation of digits and the background color of images. Then they make efforts to identify environment-independent features by measuring the distances of features across environments. However, these methods have their limitations in IoT time series prediction tasks: 1) It is difficult to split the time series data into different environments. 2) Increasing time series data from IoT devices makes the deep learning models inefficient in learning stable environment-independent features. In this paper, we propose to improve the generalizability of the regression model efficiently. On one hand, we design a regularization to select the core data from massive time series data to accelerate improving the generalizability of the model; On the other hand, a re-weighted objective function is proposed to balance the performances among time series during training instead of measuring distances of features among different environments. Experimental results on two IoT time series datasets show that our approach can effectively improve domain generalization methods and reduce around 50% training time compared to baseline methods.