ACTIVE LEARNING OF THREE-WAY DECISION BASED ON NEIGHBORHOOD ENTROPY

Addressing the issue that the entropy-based query strategy neglects the distribution of samples, a selection strategy based on neighborhood entropy is designed to combine sample distribution characteristics with entropy. The weighted entropy of all unlabeled samples in the neighborhood is employed as the measure, which can select both informative and representative unlabeled samples to label, so a more effective classifier can be trained by these samples. Furthermore, considering samples from different regions have different values, the neighborhood entropy is adopted as the decision function of the three-way decision, and an active learning of three-way decision based on neighborhood entropy is developed, namely ALTD NE. The unlabeled dataset is divided into three regions by the value of the neighborhood entropy, and then the most valuable samples in different parts are selected for labeling and employed to train the classifier, which can further improve the classification performance. Experiments are conducted on several standard classification datasets of the KEEL and UCI databases and one real classification dataset. ALTD NE has a better performance in most cases, which outperforms several comparative algorithms in ACC, F1 Value, and AUC.