Adaptive Bi-Weighting Toward Automatic Initialization and Model Selection for HMM-Based Hybrid Meta-Clustering...

Area of Science:

• Information Science
• Data Mining
• Machine Learning

Background:

• Temporal data clustering is crucial for uncovering intrinsic structures in time series and sequential data.
• Existing methods face challenges with initialization and model selection for temporal data.
• Hidden Markov Models (HMMs) offer a probabilistic framework for sequential data analysis.

Purpose of the Study:

• To propose a novel HMM-based hybrid meta-clustering ensemble to address initialization and model selection issues in temporal data clustering.
• To enhance ensemble performance through an adaptive bi-weighting scheme for optimizing consensus functions.
• To enable automatic and adaptive determination of the number of clusters.

Main Methods:

• A hybrid meta-clustering ensemble using HMM-based K-models with varying initializations.
• A bi-weighting scheme to adaptively fuse partitions from multiple HMM-based models.
• Consensus functions combined with a normalized mutual information objective function for optimal partition selection.
• Refinement using HMM-based agglomerative clustering and dendrogram-based similarity partitioning.

Main Results:

• The proposed bi-weighting scheme effectively optimizes the fusion of consensus functions.
• An optimal consensus partition is selected using a normalized mutual information-based objective function.
• The final clustering approach automatically determines the number of clusters.
• Extensive experiments show superior performance on synthetic, time series, and real-world motion trajectory datasets compared to benchmarks.

Conclusions:

• The novel HMM-based meta-clustering ensemble with a bi-weighting scheme effectively handles temporal data clustering challenges.
• The method provides robust and adaptive clustering, automatically determining the number of clusters.
• This approach demonstrates significant potential for developing improved clustering tools in information analysis and management.