Multi-relational Classification: An Ensemble Learning Approach Based On Multiple Views

In traditional data mining methods, the representation of knowledge is mainly formal system of propositional logic and the patterns can only be discovered from a single relation. But now the knoweledge in real relational database is stored in multiple relational tables and many of the complex patterns is difficult to be expressed by using the language of propositional logic.Multi-relational data mining method can discovery patterns from a relational database involves a number of complex models; furthermore, it can analyze data directly from a number of relations without converting data to a single relational table. Multi-relational data mining thus have been extensively used and developed. Multi-relational data mining mainly draws on ILP techniques and machine learning methods, is one of the important areas of data mining.This paper studies multi-relational classification. Multi-view approach can easily discovery useful patterns from relational database by directly using traditional classification methods, and then combines the classification result of each view to achieve multi-relational classification. This approach has lower complexity in algorithm than the other existing relationship classification methods, and achieves better results in classification accuracy.The combination method of multi-view approach employs the same level of integration methods, which is equivalent to equate the contribution of each view. Although this method achieves good results, there is room for improvement. This paper argues that the contributions of the various views on the classification are different because that the data sets are different, and there is complementarity between views. Using this complementary feature effectively can further improve the integration of the results. In this paper, we research and validate view complementary as the guidance to choose views to combine, and then establish the hierarchical structure of the model through heuristic search to construct a tree of multi-view.The combination of multi-view may cause that some samples are very hard to be learned. Inspired by Boosting, an ensemble learning method, we can reduce the difficulty learning the samples by boost the weight of the samples to interfere the learner. Through this way we can improve the multi-view tree method.Multi-view tree method chooses views to improve the performance of the model through heuristic search, avoiding global search, and improves performance of the combination by ensemble learning. The experimental results show that multi-view tree method is much better than the existing well-known classification systems in terms of accuracy and efficiency.