Learning Hyper-Parameters of Image Transformations for Time Series Classification

Time Series Classification (TSC) is the task of using a time series to predict the value of a categorical variable-the class of the time series. An interesting approach that has been studied in the TSC literature is converting a given time series into an image, through one of various proposed methods, and classifying the resultant image through a vision model. Furthermore, existing image transformation methods may require the values of one or more parameters. Modifying these parameter values can greatly alter the generated image, thereby impacting the performance of a vision-based classifier. However, guidelines for selecting appropriate parameter values have not been detailed. In fact, due to the diversity of time series used for TSC tasks, establishing such guidelines in a way that ensures optimal results may be impossible. Following this realization, we attempt to build Deep Learning models capable of independently learning appropriate parameter values to maximize classification accuracy. We propose a method wherein these parameters are dynamically determined through a Policy Network, which takes a time series as input. Then, an image is obtained using the resultant parameters and a pre-determined image transformation function. Finally, the image is fed into a vision model to obtain the class of the time series. This approach was tested for three pre-existing image transformation functions on the entirety of the UCR Archive of 128 TSC datasets. The results demonstrate that it may be more effective to allow models to learn the parameters of image transformation functions instead of setting them beforehand.