Time Series Anomaly Detection using Reconstruction and RBF Similarity Scores

Ramin Ghorbani, Marcel Reinders, David Tax

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Anomaly detection, Time series data, Radial Basis Function (RBF), Similarity score, Reconstruction error, Representation learning, Deep learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Anomaly detection in time series data is pivotal across various domains. The inherent challenge of scarce labeled data for anomaly detection has increased the attention toward unsupervised learning methods, in particular autoencoders and variations thereof. While these unsupervised approaches have shown promise, those that solely rely on reconstruction error often miss subtle anomalies, especially in high-dimensional or multivariate datasets. Motivated by this challenge, we introduce a novel approach that utilizes a layer of Radial Basis Function (RBF) neurons within the deep learning architectures. This RBF layer fits a nonparametric density in the hidden representation. When the neural network is trained on (predominantly) normal data, then a high RBF output indicates a high density, which in turn implies a high similarity with the normal data. Combining the RBF similarity score with the reconstruction error results in a unique anomaly score that we named the SimRec score. While our method can be adapted to a wide range of architectures, we focus on LSTM and Transformer models. We evaluate our approach on three real-world benchmark datasets, with results indicating significant improvements over the baselines. Our findings underscore the potential of the SimRec score in capturing subtle anomalies that might be overlooked by scores based on reconstruction error alone, offering a more robust and comprehensive solution for anomaly detection in time series data.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3783