V. Álvarez, S. Mazuelas, J. A. Lozano
The statistical characteristics of sequential data often change with time in practical scenarios.
Conventional supervised classification techniques adapt to such changes accounting
for a global rate of change by means of a carefully chosen parameter. However, time changes in common scenarios cannot be grasped considering only a rate of change, e.g., different data characteristics often change in a different manner. This work presents adaptive minimax risk classifiers (AMRCs) that sequentially learn classification rules with the smallest worst-case error probability. AMRCs utilize a sequence of uncertainty sets that include the varying underlying distributions with high probability and account for multidimensional and higher order changes in data characteristics. In addition, AMRCs provide performance guarantees in terms of bounds for accumulated mistakes and for error probabilities. Experiments on benchmark datasets show the improvement of AMRCs compared to the state-of-the-art.
Keywords: Concept drift, Multidimensional adaptation, Minimax Classification, Performance Guarantees
Scheduled
Data Analysis
November 7, 2023 6:40 PM
HC2: Canónigos Room 2