Random forest induction is an ensemble method that uses a random subset of features to build each node in a decision tree. The method has been shown to work well when many features are available. This certainly is the case in relational learning, especially when aggregate functions, combined with selection conditions on the set to be aggregated, are included in the feature space. This paper presents an initial exploration of the use of random forests in a relational context. We experimentally validated our approach both in a business domain, and on a structurally complex data set.
Published: A. Van Assche, C. Vens, H. Blockeel, en S. Dzeroski, A random forest approach to relational learning, ICML 2004 Workshop on Statistical Relational Learning and its Connections to Other Fields (Dietterich, T. and Getoor, L. and Murphy, K., eds.), pp. 110-116, 2004
Source: PDF, 143246 bytes