Detecting Views for Co-training with Genetic Algorithm

摘要

Co-training can improve the performance of a learner by combining labeled and unlabeled data, and the co-training algorithm outperforms other algorithms without using a feature split on some datasets according to some researches. The above datasets not only include datasets with a natural separation of their features into two disjoint sets, but also some high-dimension datasets with a random feature split. In this paper, view validation and view detection with genetic algorithm were explored. Given a feature split, the former can test whether the learner modeled by the two views will outperform a traditional leaner with just single view. The latter is related to an improved approach based on a genetic algorithm for detecting a good feature split for co-training. Some experimental results indicated that genetic algorithm can always outperform the progressive method in supplying the feature division but running slowly.