Predictive modular neural networks for unsupervised segmentation of switching time series: the data allocation...

Area of Science:

• Machine Learning
• Time Series Analysis
• Statistical Modeling

Background:

• Online unsupervised learning addresses data streams without pre-labeled examples.
• Switching time series are generated by multiple, sequentially activated sources.
• Existing methods often require distinct source separation and model development stages.

Purpose of the Study:

• To develop an integrated methodology for online unsupervised learning of switching time series.
• To introduce a general data allocation (DA) methodology combining data separation and model development.
• To analyze the theoretical properties and practical performance of the proposed DA approach.

Main Methods:

• A two-stage approach involving data assignment and model development.
• An iterative data allocation (DA) scheme where models compete for incoming data.
• Two DA modes: parallel DA (lowest prediction error) and serial DA (error below threshold).

Main Results:

• Sufficient conditions for asymptotically correct data allocation are presented.
• Numerical experiments demonstrate the effectiveness of the DA methodology.
• The iterative scheme successfully refines models using assigned data.

Conclusions:

• The proposed data allocation methodology offers an effective solution for online unsupervised learning of switching time series.
• The parallel and serial DA modes provide distinct strategies for data assignment.
• Theoretical analysis and empirical results support the validity of the approach.