Vortrag: „Time Series Change Point Detection: Adaptive LSTM-Autoencoders In Comparison To Related State-Of-The-Art Algorithms”

Change point detection (CPD) in time series data is a relevant task in modern data science fields. CPD is used to detect abrupt changes in the given data distribution, therefore having applications in a wide range of disciplines. In this talk, an approach to online change point detection is presented using unsupervised deep learning techniques. Following state-of-the-art research, LSTM-autoencoder based neural networks are exploited to perform online detection. The model adapts to new training samples to yield better prediction quality. With its flexible nature, this approach fits well to multi-dimensional time series and big data problems. This method of CPD, the so-called Adaptive LSTM-Autoencoder Change-Point Detection (ALACPD), is compared to a set of offline and online CPD algorithms in terms of performance as well as accuracy of the detected change points. The comparison between ALACPD with kernelbased change point detection (KCPD) and Bayesian online change point detect (BOCPD) yields the strengths of the individual algorithms and shows that a neural network approach provides a flexible solution to the change point detection problem. It even outperforms state-of-the-art algorithms in several data situations.