A Novel Intelligent Classifier Based on Instantiated Weighted One-dependence Estimators

Bayesian network classifiers (BNCs) are learned from data using various independence assumptions that reduce topology complexity and improve generalization, while in practice these assumptions are often described by probability-theoretic metrics. As a state-of-the-art BNC, averaged one-dependence estimators (AODE) exhibit excellent training efficiency and superior classification performance because they encompass 1-dependence relationships fully. However, the dependence relationships are complex. In this paper, by integrating Bayesian inference with independence analysis, we propose the instantiated weighted AODE (IWAODE) that can measure the dependency relationship between specific attribute values, and then derive the decomposition of the local log-likelihood for fine-tuning the weights of the committee members in AODE. Extensive evaluations were conducted on 12 datasets. Our IWAODE outperforms five leading-edge benchmarks regarding zero-to-one loss, as demonstrated by the findings.