A Relational Learning Algorithm Combining Relational Tri-training and Relational Instance-based Learning

摘要

Relational tri-training (R-tri-training for short) is a relational semi-supervised learning algorithm, it can effectively exploit unlabeled examples to learn first-order rule. However, the performance of R-tri-training is usually not stable for the unlabeled examples often are wrongly labeled during the iterative learning process. In this paper, a new relational tri-training algorithm named RIBL-R-tri-training is proposed. A new process named re-classifying is added to the iteration process of RIBL-R-tri-training, then the iteration process of RIBL-R-tri-training consists of co-labeling, re-classifying and re-training. The process of co-labeling and re-training of RIBL-R-tri-training is same as that of R-tri-training. In the process of re-classifying, for any example x in the newly labeled examples (unlabeled examples labeled by co-labeling), we use the new training examples (the newly labeled examples and original labeled examples) except x as a training examples to train RIBL (Relational instance-based learning) algorithm, then re-classify the example x. For RIBL is robust with respect to noise in the examples or missing attribute values, the new training examples after re-classifying has little noise than that before re-classifying. Experiments on well-known benchmarks in chemoinformatics show that RIBL-R-tri-training can more effectively enhance the performance of the hypothesis learned than R-Tri-training.