Ayla Jungbluth and Johannes Lederer’s contribution to the Discussion of ‘the Discussion Meeting on Probabilistic and statistical aspects of machine learning’
本文评论了Li等人关于用深度学习自动检测时间序列变点的论文，并介绍了作者自己的BatchCP方法，该方法通过避开变点选择批次来改进深度学习预测，可与所评论文的方法结合使用。
We have read the paper ‘Automatic Change-Point Detection in Time Series Via Deep Learning’ by Jie Li, Paul Fearnhead, Piotr Fryzlewicz, and Tengyao Wang with great interest. We congratulate the authors for an important and forward-looking contribution. We agree that time-series data often have change points that can degrade the prediction quality considerably. The paper’s use of deep learning to detect change points is, therefore, definitely interesting. Besides, our own work shows that change-point detection is also vital for time-series pipelines that are based on deep learning themselves (Jungbluth & Lederer, 2023). We have also shown that change points can be accounted in deep learning in a surprisingly simple and effective way. Indeed, most deep-learning-based forecasting methods learn by taking batches from the time-series data. In essence, our BatchCP method selects batches from the time series such that change points are avoided while we are still able to use the entire time series as data. For this, BatchCP needs to know where the change points are before the training begins. Based on this information, the batches are then deemed suitable or not as illustrated in Figure 1. Visualization of the change-point method BatchCP, taken from the mentioned reference. It shows an example of a batch selection around a change-point. The points represent time-series data. The first area shows a batch created by choosing t = 1 as the start index, and with a selected batch size of s = 6, the batch contains the points t = 1 to t = 6. The change-point is located at t = 7 and is therefore outside the detected batch. The batch is therefore permissible. The area in the middle shows a prohibited batch. The start index of this batch is t = 4 and the end index at t = 9, so the change-point is located inside the batch and is not allowed in training. A new batch must be found. The area to the right indicates another permissible batch, since the change-point lies outside of it. BatchCP is agnostic to the way change points are detected and, therefore, should work very well with the method proposed in the discussed paper. Indeed, the combination of BatchCP and the proposed method in their paper would equip modern architectures such as DeepAR and transformers with a quite automated approach to deal with change points. Our numerical studies still need to be completed, but the initial results look promising already.