Bayesian network based label correlation analysis for multi-label classifier chain

Classifier chain (CC) is a multi-label learning approach that constructs sequence of binary classifiers according to label order. Each classifier in the responsible for predicting relevance one label. When training label, proceeding labels will be taken as extended features. If features are highly correlated performance improved, otherwise, not influenced or even degraded. How discover correlation and determine order critical CC approach. This paper employs Bayesian network (BN) model correlations proposes new BN-based method (BNCC). Conditional entropy used describe dependency relations among labels, BN built up by taking nodes weights edges their relations. A scoring function proposed evaluate structure, heuristic algorithm introduced optimize BN. At last, applying topological sorting on optimized BN, constructing derived. Experiments demonstrate feasibility effectiveness method.