CNN Approaches to Classify Multivariate Time Series Using Class-specific Features

Many smart data services (e.g., smart energy, smart homes) collect and utilize time series data (e.g., energy production and consumption, human body movement) to conduct data analysis. Among such analysis tasks, classification is a widely utilized technique to provide data-driven solutions. Most existing classification methods extract a single set of features from the data and use this feature set for classification across multiple classes. This often ignores the reality that different and class-specific subsets of the initial feature set may better facilitate classification. In this paper, we propose two convolutional neural network (CNN) models using class-specific variables to solve the multi-class classification problem over multivariate time series (MTS) data. A new loss function is introduced for training the CNN models. We compare our proposed methods with 13 baseline approaches using 14 real datasets. The extensive experimental results show that our new approaches can not only outperform other methods on classification accuracy, but also successfully identify important class-specific variables.