An end-to-end approach for benchmarking time-series models using autoencoders

Abstract

Data generation and augmentation have become an important process in the area of model training and validation by allowing us to measure a model’s robustness in comparison with others. Training on sparse and small datasets can reduce the effectiveness of a model as well as make it prone to overfitting. Thus, it is important to be able to validate its accuracy on data that has the same inherent characteristics as the original but includes subtle variances potentially present in real-world data. Our novel contribution makes use of a generative modeling technique to encode such discriminative features and learn their representation through the use of an encoder and generate new data through the decoding of such encodings. We propose to train the decoder through the use of a differentiable version of the popular Dynamic Time Warping (DTW) algorithm to formulate the objective loss of our end-to-end data generation model. The model will be able to generate suitable validation data for testing existing datasets, especially low resource and scarce ones.