K.-Y. Whang, J. Jeon, K. Shim, J. Srivatava (Eds.): PAKDD 2003, LNAI 2637, pp. 595–601, 2003. © Springer-Verlag Berlin Heidelberg 2003 Upgrading ILP Rules to First-Order Bayesian Networks Ratthachat Chatpatanasiri and Boonserm Kijsirikul Department of Computer Engineering, Chulalongkorn University, Pathumwan, Bangkok, 10330, Thailand. ratthachat.c@student.chula.ac.th, boonserm.k@chula.ac.th Abstract. Inductive Logic Programming (ILP) is an efficient technique for re- lational data mining, but when ILP is applied in imperfect domains, the rules in- duced by ILP often struggle with the overfitting problem. This paper proposes a method to learn first-order Bayesian network (FOBN) which can handle imper- fect data powerfully. Due to a high computation cost for directly learning FOBN, we adapt an ILP and a Bayesian network learner to construct FOBN. We propose a feature extraction algorithm to generate features from ILP rules, and use these features as the main structure of the FOBN. We also propose a propositionalisation algorithm for translating the original data into the single ta- ble format to learn the remaining parts of the FOBN structure and its conditional probability tables by a standard Bayesian network learner. 1 Introduction Inductive Logic Programming (ILP) plays the central role to relational data mining, but the first-order rules induced by ILP often struggle with the overfitting problem (see [10]). Recently, a very nice combination between a first-order and a graphical model is proposed by Koller et al. [6] and by Kersting and De Raedt [7] in the name of prob- abilistic relational model (PRM) and Bayesian logic program (BLP), respectively. In this paper, we will call both models as first-order Bayesian network (FOBN). The model of FOBN is very powerful to handle noisy data by the power of probabilistic theories, and also an expressive model by the power of first-order model combining with Bayesian network. However, the algorithms to learn FOBN proposed in [6] and [7] centrally concern in discovery tasks while the algorithm to learn FOBN as a classi- fier has not been well discussed. Hence, this paper proposes an efficient framework to learn FOBN as classifier. 2 The Framework Overview Botta et al. [2] have recently shown that the relational learning problem is linked to the exponential complexity, and Chickering [3] has also shown that learning a Bayesian