Analysis of Time Series Anomalies Using Causal InfoGAN and Its Application to Biological Data

Takaya Ueda, Masataka Seo, Yukako Tohsato, Ikuko Nishikawa

Abstract: Data generation is an inverse function of the data recognition, which is a map from a low dimensional latent space to a high dimensional data. Generator of normal data acquires the data distribution in its latent space, and is expected to generate any normal data, but unable to generate abnormal data. Therefore, the generator can be used for the anomaly detection. We propose a model of Generative Adversarial Nets (GANs) with an encoder to characterize the data anomaly. GANs are trained by normal data, and the encoder is trained to infer the latent space of the trained generator. Then, by combining the encoder and the generator, anomaly of the data input to the encoder is characterized by the data reconstruction error and by its latent representation. If any interpretable representation is obtained for the latent space, we can characterize the abnormality by the latent variable. In this study, Causal InfoGAN is trained by normal time series data to acquire temporal state variables in the latent space. Then, a temporal abnormality can be characterized by an abnormal transition in the state space. Proposed method is applied to a simple toy data, which mimics a cell division in the biological system, and it is observed that the state transition which expresses the normal time development is obtained in the latent space.