Time Series Representation for Real-World Applications of Deep Neural Networks

Neural networks have a proven usefulness at predicting, denoising or classifying time series. However, the performance of deep learning models is bound to the size of the input window. Yet, no common method has emerged to determine the optimal window size. In this paper, we compare two heuristics and three event detection algorithms to find the best time representation for three different tasks, using one simulated and two real-world datasets. The two real-world applications are the electricity disaggregation for energy efficiency in buildings and the detection of fibrillation for diagnosis in cardiology. We compare the obtained window sizes with the experimental values from previous research and we experimentally validate the relevance of the results using both convolutional and recurrent deep neural networks. Results confirm the impact of the sequence length on model performance and show that window sizes cannot be simply transferred to another dataset, even for the same problem. We also find that the false nearest neighbors method can reliably estimate the window size and can help with the tedious work of finding the right time representation.