Reconstruction-based unsupervised drift detection over multivariate streaming data

In real-world applications, data streams are gener-ated all the time. Real-time data processing of complex multi-variate data becomes essential for many downstream analysis tasks. However, real-world data is not bound to be of the same distribution - the environment where it is recorded could be rapidly evolving, making it a challenge to apply a stationary model throughout the whole flow of data. Moreover, labels are often expensive to acquire or delayed in such scenarios. This paper considers the severe problem in an unsupervised setting, where we detect the distributional drifts in the input data stream without considering the data labels or specific classifiers. We propose AECDD (Autoencoder-based Concept Drift Detec-tor), a reconstruction-based unsupervised drift detection model using an Autoencoder to track changes in data distribution. More specifically, instead of detecting drifts by tracking the classifi-cation error change as in many existing approaches, we track the reconstruction error of the Autoencoder in an unsupervised manner. Our empirical evaluation shows that AECDD captures the drifts well in multivariate data streams. Finally, we also demonstrate the drift in the reconstruction error space by an intuitive visualization.