Improving imbalanced classification using near-miss instances

The class imbalance is a major issue in classification, i.e., the sample size of rare (positive) often performance bottleneck. In real-world situations, however, “near-miss” positive instances, negative but nearly-positive are sometimes plentiful. For example, natural disasters such as floods rare, while there relatively plentiful near-miss cases where actual did not occur water level approached bank height. We show that even when true quite limited, disaster forecasting, accuracy can be improved by obtaining refined label-like side-information “positivity” (e.g., river) to distinguish from other negatives. Conventional cost-sensitive classification cannot utilize side-information, and small causes high estimation variance. Our approach line with learning using privileged information (LUPI), which exploits for training without predicting itself. theoretically prove our method reduces variance, provided instances plentiful, exchange additional bias. Results extensive experiments demonstrate tends outperform or compares favorably existing approaches.