Report of ILP 2010

Francesca A. Lisi, Università degli Studi di Bari “A. Moro”, Italy

The 20th International Conference on Inductive Logic Programming (ILP 2010) has been held in Firenze, Italy, chaired by Paolo Frasconi and Francesca A. Lisi. The conference was well attended with a high level of participation and discussion. Thanks to the generous support of the ALP, discounted registration fees were granted to 6 students having a paper accepted at the conference and a well motivated financial need.

The program featured 31 accepted papers (16 for oral presentation and 15 for poster presentation) out of 44 submissions, 3 invited talks, and 2 tutorials. The conference proceedings will be made available as a volume of the Springer LNAI series. A selection of the accepted papers will appear in a special issue of the Machine Learning Journal.

The contributions to ILP 2010 span from theory to practice, mainly within the stream of the so-called Statistical Relational Learning  (SRL). Here is a summary of the principal new ideas:

·         Best Student Paper (award sponsored by the Machine Learning journal): “From inverse entailment to inverse subsumption” (Yamamoto, Inoue, Iwanuma) shows how IE can be reduced to inverse subsumption by preserving its completeness.

·         “Approximate Bayesian computation for the parameters of PRISM programs” (Cussens) presents a Bayesian method for approximating the posterior distribution over PRISM parameters.

·         “Applying the Information Bottleneck Approach to SRL: Learning LPAD Parameters” (Riguzzi, Di Mauro) adopts the Information Bottleneck approach for learning LPAD parameters.

·         “Extending ProbLog with Continuous Distributions” (Gutmann, Jaeger, De Raedt) extends ProbLog with abilities to specify and infer over continuous distributions.

·         “Probabilistic Rule Learning” (De Raedt, Thon) upgrades rule learning to a setting in which both the examples and their classification can be probabilistic.

·         “Boosting Relational Dependency Networks” (Natarajan et al.) proposes the use of gradient tree boosting in RDNs.

·         “Multitask Kernel-based Learning with First-Order Logic Constraints” (Diligenti et al.) defines a general framework to integrate supervised/unsupervised examples with background knowledge in the form of FOL clauses into kernel machines.

·         “Stochastic Refinement” (Tamaddoni-Nezhad, Muggleton) introduces the notions of stochastic refinement operator and search.

·         “Hypothesizing about Networks in Meta-level Abduction” (Inoue, Doncescu, Nabeshima) deals with completing causal networks by means of meta-level abduction.

·         “Learning discriminant rules as a minimal saturation search” (Lopez, Martin, Vrain) defines a non-blind bottom-up search strategy for hypotheses.

·         “Speeding up Planning through Minimal Generalizations of Partially Ordered Plans” (Cernoch, Zelezny) presents an ILP framework for planning which exploits existing plans in new similar planning tasks.

·         “Exact Data Parallel Computation for Very Large ILP Datasets” (Srinivasan, Faruquie, Joshi) shows how distributed computing can be used effectively in ILP.

·         “Automating the ILP Setup Task: Converting User Advice about Specific Examples into General Background Knowledge” (Walker et al.) introduces some techniques to automate the use of ILP systems for a non-ILP expert.

·         “Fire! Firing Inductive Rules from Economic Geography for Fire Risk Detection” (Vaz, Santos Costa, Ferreira) provides an elegant and powerful approach to spatial data mining by coupling Spatial-Yap with an ILP engine.

·         “Multivariate Prediction for Learning on Semantic Web” (Huang, Tresp, et al.) proposes multivariate analysis methods for the prediction of potential relationships and attributes in Semantic Web data.

The invited talks were very inspiring. Michael Kifer argued that RIF, a W3C recommendation for the exchange of Semantic Web rules, is a major opportunity to rekindle the interest in logic programming. Avi Pfeffer presented a new probabilistic programming language named Figaro that is designed with practicality and usability in mind. David Poole provided a new vision of the future Web as a World Wide Mind where the so-called Semantic Science will play a key role and claimed that SRL and ILP need to be a foundation of the Semantic Web.

The tutorials concerned topics of relevant interest to ILP researchers. Francesco Scarcello and Gianluigi Greco illustrated the main structural decomposition methods proposed in the literature of database and graph theory to identify easy instances of hard problems. Volker Tresp described some of the main approaches to multivariate modeling in Statistics and shown its application to Relational Learning.

In order to celebrate the 20th birthday of the conference series, a panel discussion with title “ILP turns 20” was organized where the invited experts Ivan Bratko, Luc De Raedt, Peter Flach, Katsumi Inoue, Stephen Muggleton, David Poole and Ashwin Srinivasan expressed their own view of the main achievements of ILP and its contributions to the broader areas of machine learning, logic programming, and artificial intelligence in these 20 years, as well as their own view on current and future trends. Most of the experts agreed on the fact that predicate invention has been the major contribution of ILP to the abovementioned fields and should be further investigated in the next future.