Cost Sensitive Time-Series Classification

Introduction: One of the key sources of performance degradation in the field of time-series classification is the class imbalance problem. In real-world datasets, the minority class (Positive class) is outnumbered by abundant majority (negative) class instances. Objective: Develop a cost-sensitive time-series classification framework. Challenge 1: Minimum classification error criterion (e.g 0-1 loss function) based classification models generate biased models towards the majority class causing higher misclassification error for minority class examples (important class). Solution: Use differentially weighted loss function having variable misclassification cost for false positive and false negative errors. Challenge 2: Predetermination of misclassification cost values from domain experts. Solution: Estimating misclassification cost from data. Challenge 3: Interpretable Model. Solution: Learning interpretable temporal patterns (shapelets) for time-series classification.