Predicting Future Classifiers for Evolving Non-linear Decision Boundaries

Abstract

In streaming data applications, the underlying concept often changes with time which necessitates the update of employed classifiers. Most approaches in the literature utilize the arriving labeled data to continually update the classifier system. However, it is often difficult/expensive to continuously receive the labels for the arriving data. Moreover, in domains such as embedded sensing, resource-aware classifiers that do not update frequently are needed. To tackle these issues, recent works have proposed to predict classifiers at a future time instance by additionally learning the dynamics of changing classifier weights during the initial training phase. This strategy bypasses the need to retrain/relearn the classifiers, and thus the additional labeled data is no longer required. However, the current progress is limited to the prediction of linear classifiers. As a step forward, in this work, we propose a probabilistic model for predicting future non-linear classifiers given time-stamped labeled data. We develop a variational inference based learning algorithm and demonstrate the effectiveness of our approach through experiments using synthetic and real-world datasets.

D.Dhaka—Contributed while working at NEC Corporation.