Minimax Classification with Multidimensional Adaptation to Time Changes
Verónica Álvarez
Abstract:
In common practical scenarios, the characteristics of sequential data often change with time and different data characteristics often change in a different manner. Current supervised classification techniques adapt to time changes accounting for a global rate of change by means of a carefully chosen learning rate, forgetting factor, or weight factor. However, changes in different data characteristics cannot be grasped considering only a rate of change. This work presents adaptive minimax risk classifiers (AMRCs) that sequentially learn classification rules accounting for multidimensional and high-order time changes in data characteristics. In addition, AMRCs provide performance guarantees in terms of bounds for instantaneous error probabilities and for accumulated mistakes. The numerical results show the reliability of the presented performance guarantees and the classification improvement of AMRCs compared to the state-of-the-art.