Multimodal Batch-Wise Change Detection.

Area of Science:

• Machine Learning
• Data Science
• Statistical Modeling

Background:

• Existing change detection (CD) algorithms struggle with batch-wise multimodal data, showing low detection power or poor false positive control.
• Current methods often assume a single distribution for stationary conditions, which is inadequate for multimodal scenarios.

Purpose of the Study:

• To develop a novel change detection algorithm, MultiModal QuantTree (MMQT), for batch-wise and multimodal data.
• To address the limitations of existing CD algorithms in handling multiple distributions within stationary conditions.

Main Methods:

• MMQT utilizes a single histogram to model batch-wise multimodal stationary conditions.
• The algorithm automatically identifies the modality of incoming batches and employs modality-specific statistics for change detection.
• Theoretical properties of QuantTree are leveraged for automatic modality number estimation and principled false-positive control calibration.

Main Results:

• MMQT demonstrates high detection power and accurate false-positive control in both synthetic and real-world multimodal CD problems.
• Experiments validate the algorithm's effectiveness in stream learning applications, including detecting concept drift and novel class emergence.

Conclusions:

• MMQT offers a robust solution for change detection in complex, multimodal, batch-wise data streams.
• The algorithm shows significant promise for applications in stream learning and monitoring input distribution changes.