Multi-relational Classification Based On Decision Tree Algorithm

Multi-relational data mining is one of the rapidly developing subfields of data mining. The classical data mining approaches can only look for patterns in single relation, and it is difficult to look for complex relational patterns which involved in multi-relational databases. But in real world, the organizations of data usually use the multi-relational models in relational databases. So the classical data mining approaches are not suit for this situation. That is why this paper chooses a multi-relational data mining algorithm as our research object.Efficiency and scalability have always been important concerns in the field of data mining. The complication of multi-relational data mining has made higher demands, because of the larger hypothesis spaces and the more complex evaluation methods. So we focus on the improving of efficiency and scalability of the multi-relational data mining algorithm.The main bottleneck of improving efficiency of the algorithm is the size of the hypothesis space. Therefore, the key to improve efficiency is decreasing the size. In this paper we talk about multi-relational classification approach based on the decision tree algorithm. We improve the efficiency of the algorithm on two aspects as follows.First of all, we use a sufficient table to reduce the number of tables which will be joined in each hypothesis model in the evaluation process. In addition, the target tuple ID propagation approach is used to save the connection information, thus the connections will not be wasted even if the hypothesis model is not optimal.Finally, in this paper proposes a new multi-relational classification algorithm and experiments on the PKDD CUP'99 financial data set. The experimental results show that our algorithm improves the efficiency of MRDTL obviously.