Long-Tailed Time Series Classification with Noisy Labels

Existing deep time series classification methods typically rely on a strict assumption that data is evenly distributed across classes with error-free labels. However, in practical applications, time series data faces a joint dilemma of long-tailed distributions and noisy labels, which leads to performance degradation of these methods. To this end, a novel Cluster-oriented Multi-dimensional Distillation method with adaptive Decision perception (CMDD) is proposed via integrating pure graph-driven semantic extraction, cluster-oriented multi-dimensional distillation, and long-tailed class-aware optimization into a unified architecture, which effectively achieves the classification of long-tailed time series data with noisy labels via joint learning of semantic representations, data distillation, and decision optimization. Meanwhile, extensive experiments are conducted on four long-tailed time series datasets with various noise ratios, and the results demonstrate the superiority of CMDD. Our code is available at https://github.com/qinqinhan/CMDD.