Linear dependency modeling for classifier fusion and feature combination.

Area of Science:

• Machine Learning
• Data Fusion
• Pattern Recognition

Background:

• Traditional data fusion methods often rely on the restrictive independent assumption.
• Existing dependency modeling techniques require specific classifier distributions or complex joint density estimations.

Purpose of the Study:

• To propose a new framework for modeling feature dependency in data fusion without assuming feature or classifier distributions.
• To overcome the challenges associated with estimating high-dimensional joint densities.

Main Methods:

• Developed a framework proving feature dependency can be modeled via a linear combination of posterior probabilities.
• Introduced Linear Classifier Dependency Modeling (LCDM) for classifier-level fusion.
• Introduced Linear Feature Dependency Modeling (LFDM) for feature-level fusion.
• Optimized models by maximizing the margin between genuine and imposter posterior probabilities.

Main Results:

• LCDM and LFDM demonstrated superior performance over existing methods under non-normal distributions.
• LFDM outperformed all other methods, including classifier-level fusion, on real datasets.
• Experimental validation was conducted using both synthetic and real-world datasets.

Conclusions:

• The proposed framework effectively models feature dependency without restrictive distribution assumptions.
• LFDM offers the best performance for data fusion tasks, particularly in complex, real-world scenarios.