Optimizing Detection Time and Specificity: Early Classification of Time Series with Sensitivity Constraint

Abstract

From the perspective of sequential decision making, we propose a novel approach for early classification of time series under the Neyman--Pearson paradigm that incorporates a sensitivity constraint. We explicitly formulate the optimal solution, which can be practically obtained utilizing plug-in estimators such as recurrent neural networks. Cast as a constrained multi-objective optimization problem, we establish the Pareto optimality balancing earliness and classification accuracy. Our approach visualizes the inherent trade-off between earliness and specificity, ensuring informed decision making without compromising sensitivity. Experimental validation confirms the feasibility of our approach, demonstrating its potential in various real-world applications.