IF2CNN: Towards non-stationary time series feature extraction by integrating iterative filtering and convolutional neural networks

Abstract

Time series is a common data type that appears in various fields. However, time series processing is still considered one of the most challenging problems in data mining because of its unique properties, such as noise and non-stationarity. In an effort to overcome these limitations, we present a framework, called IF2CNN, that integrates the iterative filtering (IF) method and convolutional neural networks (CNNs) for automatic feature learning for time series. First, IF is leveraged to decompose the raw non-stationary time series into intrinsic mode functions (IMFs), which are then converted into image format data. Second, CNN is designed to automatically learn features from the image format data, which can help to extract deep and global features of the time series. Besides, the use of IF and CNN technologies makes the proposed framework not only have the advantage of dealing with the non-stationarity of the time series, but also provides a good generalization ability for small training datasets. To evaluate the performance of IF2CNN, two different strategies are used to test the role of the derived features of CNN (called CNN features). The first strategy computes the feature importance through the methods based on decision trees, such as gradient boosting decision trees and random forest, and the other one tests the validity of the derived features by performing specific prediction tasks. Furthermore, three real datasets from different fields are used in our experiments. The results show that the CNN features have an overwhelming advantage in feature importance and significant improvements in prediction tasks.