DGMSCL: A dynamic graph mixed supervised contrastive learning approach for class imbalanced multivariate time series...

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Imbalanced Multivariate Time Series Classification (ImMTSC) is crucial for identifying rare but significant events like system faults or medical anomalies.
• Minority-class instances in ImMTSC are challenging due to their rarity, randomness, and complex spatial-temporal dependencies, often leading to classification interference.
• Existing contrastive learning methods struggle to aggregate features from neighboring minority instances, hindering effective representation in imbalanced datasets.

Purpose of the Study:

• To propose a novel dynamic graph-based mixed supervised contrastive learning method (DGMSCL) for ImMTSC.
• To enhance the representation of minority-class features without increasing sample size and improve their separation from other instances.
• To achieve superior classification performance on imbalanced time series data.

Main Methods:

• Reconstruction of input sequences into dynamic graphs.
• Application of a hierarchical attention graph neural network (HAGNN) for discriminative instance embedding.
• Introduction of a mixed contrast loss incorporating weight-augmented inter-graph supervised contrast (WAIGC) and context-based minority class-aware contrast (MCAC).
• Adjustment of sample weights to prioritize minority-class learning and enhance gradient gains.

Main Results:

• DGMSCL consistently outperforms existing baseline models across various imbalanced scenarios and datasets.
• Significant improvements observed in overall classification accuracy, including higher average F1-score, G-mean, and kappa coefficient.
• Demonstrated strong generalization capabilities on a real-world power grid dataset.

Conclusions:

• The proposed DGMSCL method effectively addresses the challenges of ImMTSC by improving minority-class feature representation and separation.
• DGMSCL offers a robust solution for accurately identifying critical events in imbalanced time series data.
• The method shows significant potential for real-world applications requiring high-accuracy classification of rare events.