Research Of Label Correlation Based Multi-label Classification Algorithms In Multi-relational Data

With the rapid development of mobile internet, all kinds of mobile internet applications have become the most important information interaction platforms, these information interaction platforms form series virtual social networks. Multiple virtual social networks can be integrated as a multi-relational social network through the mapping with the nodes or relations, and we usually use t he multi-relational graph to describe the multi-relational social network. The research of multi-relational multi-label classification algorithms is essential to precise marketing, social network analysis and social information mining.In multi-label classification problems, how to effectively use the label dependency is very important to improve the performance of classification algorithms. As to the multi-relational multi-label classification problems, the label dependency includes two aspects: the label dependency in the node content attribute and that implied in the relation topology. The emphasis of this research is how to efficiently dig this two kinds of label dependency, and design more targeted multi-label classification algorithms based on that.Based on the idea of content attribute label dependency, we use the class co-occurrence information to calculate the label dependency, and propose a multi-relational multi-label classification algorithm based on it(MRML-C). This algorithm combines the label space clustering partition strategy which effectively reduces the algorithm complexity by transform the multi-label problem into several smaller sub-problems. The comparing experiment results show that the label space partition strategy effectively solves the problem of label explosion which is always happen in the multi-label problems, and the MRML-C shows better performance in most data sets.Based on the idea of relational topology label dependency, we use the class co-occurrence information and the relational topology information together to calculate the label dependency, and put forward multi-relational multi-label classification algorithm based on it(MRML-R). This algorithm firstly partition the label space according to the relational topology label dependency with clustering method, then use problem transformation method to transform the multi-label problem into single-label problem, and in the process of training models, MRML-R uses the random forest algorithm with a new bagging strategy based on the random walk which integrate the relational topology information. Finally adopts the majority voting strategy to integrate the predict results in each label subspace. The experimental results indicate that MRML-R algorithm outperforms the others.