AdOn HDP-HMM: An Adaptive Online Model for Segmentation and Classification of Sequential Data

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Automated classification of sequential data is crucial for real-time applications like activity recognition and financial analysis.
• Existing methods often require predefined class sets and struggle with evolving data distributions or continuous data streams.
• There is a need for adaptive online systems that can classify data on-the-fly without a fixed number of classes.

Purpose of the Study:

• To develop a principled solution for adaptive online classification of sequential data.
• To address the limitations of offline classification methods in dynamic and evolving data environments.
• To enable classification with an unlimited and potentially changing number of classes.

Main Methods:

• Developed an adaptive online system integrating Markov switching models and hierarchical Dirichlet process priors.
• Introduced an adaptive 'learning rate' to balance parameter retention and adaptation to new observations.
• Ensured the system meets memory and delay constraints for streaming data contexts.

Main Results:

• Demonstrated remarkable performance in segmentation and classification on synthetic and real-world video datasets.
• Showcased effectiveness in handling sequences with evolutionary distributions and previously unseen classes.
• Validated the system's capability for on-the-fly classification in streaming scenarios.

Conclusions:

• The proposed adaptive online system offers a robust solution for classifying sequential data in dynamic environments.
• The approach successfully handles an unlimited number of classes and evolving data, outperforming traditional methods.
• This work advances the field of real-time sequential data analysis, particularly for non-stationary data streams.