A Multitask Dynamic Graph Attention Autoencoder for Imbalanced Multilabel Time Series Classification.

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Multilabel time series classification is crucial for personalized predictions and risk assessments.
• Existing methods struggle to model label relevance and address imbalanced label distributions.
• Current balancing strategies often disregard label relevance, leading to information loss and bias.

Purpose of the Study:

• To propose a novel dynamic graph attention autoencoder-based multitask (DGAAE-MT) learning framework.
• To accurately model label relevance for each instance in multilabel time series.
• To improve classification accuracy for imbalanced label distributions without information loss or sampling bias.

Main Methods:

• Utilized a dynamic graph attention-based graph autoencoder to capture complex label relevance.
• Employed a dual-sampling strategy for data balancing.
• Implemented a cooperative training approach for enhanced classification performance.

Main Results:

• Achieved a mean average precision (mAP) of 0.955.
• Obtained an F1 score of 0.978 on a mixed medical time series dataset.
• Demonstrated superior performance compared to state-of-the-art methods.

Conclusions:

• The DGAAE-MT framework effectively models label relevance and addresses data imbalance in multilabel time series classification.
• The proposed method significantly enhances classification accuracy, particularly for low-frequency classes.
• DGAAE-MT offers a robust solution outperforming existing approaches.