Learning Time Series Detection Models from Temporally Imprecise Labels.

Area of Science:

• Machine Learning
• Signal Processing
• Biomedical Informatics

Background:

• Learning from time series data with precise labels is well-established.
• Temporally imprecise labels, common in mobile health (mHealth), pose a significant challenge for traditional models.
• Existing methods often struggle with noisy or misaligned event timestamps.

Purpose of the Study:

• To develop a general learning framework for time series detection using temporally imprecise labels.
• To address the challenge of noisy timestamps in event detection for applications like mHealth.
• To provide a robust method that outperforms current alternatives in handling label noise.

Main Methods:

• A novel learning framework accommodating various base classifiers and noise models was proposed.
• The framework directly models the uncertainty associated with noisy event timestamps.
• Experiments were conducted using real-world mobile health data.

Main Results:

• The proposed framework significantly outperformed baseline methods.
• Performance gains were observed compared to assuming noise-free labels and multiple instance learning approaches.
• Manual alignment of labels did not yield results as strong as the proposed method.

Conclusions:

• The developed framework offers a superior approach for learning time series detection models with temporally imprecise labels.
• This method is particularly valuable for mHealth research where precise event annotation is difficult.
• The findings suggest a new direction for robust time series analysis in the presence of label noise.