A new relational Tri-training system with adaptive data editing for inductive logic programming

摘要

Relational Tri-training (R-Tri-training for short), as a relational semi-supervised learning system, can effectively exploit unlabeled examples to improve the generalization ability. However, the R-Tri-training may also suffer from the common problem in traditional semi-supervised learning, i.e., the performance is usually not stable for the unlabeled examples often be wrongly labeled and accumulated during the iterative learning process. In this paper, a new Relational Tri-training system named ADE-R-Tri-training (R-Tri-training with Adaptive Data Editing) is proposed. Not only does it employ a specific data editing technique to identify and correct the examples possibly mislabeled throughout the co-labeling iterations, but it also takes an adaptive strategy to decide whether to trigger the editing operation according to different cases. The adaptive strategy consists of five pre-conditional theorems, all of which ensure the iterative reduction of classification error under PAC (Probably Approximately Correct) learning theory. Experiments on well-known benchmarks show that ADE-R-Tri-training can more effectively enhance the performance of the hypothesis learned than R-Tri-training.