ROBUST TIME SERIES RECOVERY AND CLASSIFICATION USING TEST-TIME NOISE SIMULATOR NETWORKS

Area of Science:

• Data Science
• Machine Learning
• Signal Processing

Background:

• Time-series data frequently suffers from corruption like missing samples and sensor noise.
• Correcting these issues is challenging when the noise model is unknown during deployment.
• This prevents the use of pre-trained classifiers on clean source data.

Purpose of the Study:

• To develop a general framework for time-series data that can estimate noise models at test time using unlabeled data.
• To enable the use of pre-trained classifiers on corrupted time-series data.
• To demonstrate a method for robust time-series analysis and prediction.

Main Methods:

• A coupled decoder model and a learned noise model simulator were employed.
• The framework estimates the noise model entirely at test time using unlabeled samples.
• The approach cleans corrupted data to match source training data statistics.

Main Results:

• The proposed framework effectively "cleans" corrupted time-series data.
• Cleaned data can be directly utilized with pre-trained classifiers for accurate predictions.
• Empirical studies across diverse domains and sensor types confirm the method's effectiveness and generality.

Conclusions:

• The developed framework offers a general solution for handling corrupted time-series data without prior knowledge of noise models.
• This approach significantly enhances the robustness of machine learning models in real-world applications.
• The method facilitates reliable time-series analysis and prediction across various sensor data types.