decisionlevel

Decisionlevel, also known as decision-level fusion, is a concept in information fusion and machine learning that involves combining independent decisions from multiple sources to produce a final decision. It sits above data-level and feature-level fusion, which merge raw data or extracted features before classification. In decisionlevel fusion, each source outputs a decision—such as a class label or a confidence score—and these outputs are merged through methods like majority voting, weighted voting, or probabilistic techniques such as Bayesian model averaging, Dempster-Shafer theory, or stacking with a meta-classifier. Outputs may be calibrated to ensure comparability of confidence across sources.

Applications of decisionlevel fusion are broad and include biometric systems that combine multiple classifiers (for example, face, fingerprint, and iris), surveillance and multimedia analysis, and remote sensing or robotics, where fusing classifier decisions can improve reliability and robustness.

Advantages include increased robustness to individual classifier errors, modularity (ability to add or remove sources without retraining all components), and potential privacy or bandwidth benefits since raw data need not be shared. Disadvantages include potential loss of information due to discarding raw data, possible redundancy when source decisions are correlated, and the need for compatible output representations and proper calibration to avoid biased fusion. The effectiveness of a decisionlevel approach depends on the diversity and quality of the contributing decisions, the fusion method used, and the problem domain.

See also: sensor fusion, ensemble methods, voting classifiers, Dempster-Shafer theory, stacking.