See also: Bled, ICML'99, ILP'99, Nicolas Lachiche

ILP'99

June 24-27, Bled, Slovenia

by Nicolas Lachiche <[email protected]>

The Ninth International Workshop on Inductive Logic programming, ILP-99, was held in Bled, Slovenia, 24-27 June 1999. It was collocated with the Sixteenth International Conference on Machine Learning, ICML-99, held 27-30 June 1999. There was an overlap day (27 June) where the following events took place: a poster session were all the papers accepted at ILP-99 and ICML-99 were presented, and a joint invited talk by J. Ross Quinlan. In addition, ILP-99 participants were allowed to attend ICML-99 sessions on that day.

Three invited talks were given at ILP-99, by Heikki Mannila, Daphne Koller and J. Ross Quinlan. Heikki Mannila showed that ILP is a strong candidate to provide inductive databases with a theoretical background. Inductive databases are databases where the user is given a view on the data plus all the generalisations -the association rules for instance- holding on the data. The associated query language allows the user to select rows satisfying A=1 and B<5 as well as association rules having C on the right-hand side. In her talk, Daphne Koller presented the probabilistic relational models. Those models combine advantages of predicate logics and Bayesian networks. Actually, they integrate uncertainty with relational models. They are certainly of high interest for the ILP community. An article presenting the learning aspect of her talk is in the proceedings. J. Ross Quinlan's talk was given to both the ICML and ILP communities. He focused on the design of classifier-learning systems. Since most of them use a propositional representation, his talk was rather oriented toward the machine learning community. Though, relational learning belongs to the most valuable assets of machine learning he listed. Moreover, the wish he made that in the future machine learning people pay more attention to interpretability opens away for the more expressive languages of ILP.

24 regular papers were presented. About half of them were about some fundamental aspects of ILP, such as refinement operators, learnability and heuristics. Nine papers focused on applications of ILP systems. Some applications considered were the automatic layout of magazines, predicting the biodegradability of chemical compounds in water, the recognition of Thai Character with an interesting combination of an ILP rule learner and a neural network, and the prediction of chemical toxicity. Five of the papers focusing on applications were about learning languages in logic (LLL), mainly on part-of-speech tagging. The remaining papers presented combinations of ILP with other machine learning techniques or new representations. All of the papers presented interesting works, but I would like to point out two papers among those I classified as presenting combinations of ILP with other machine learning techniques or representations. Both are about combining Instance-Based Learning and ILP: "Analogical Prediction" by Stephen Muggleton and Michael Bain, and "Instance Based Function Learning" by Jan Ramon and Luc De Raedt. The former focuses Progol's search on the instance to classified so it provides a kind of instance-based learning without the need of a distance between instances. Actually, analogical prediction doesn't look for any kind of neighbours, but instead it uses information in the test set to guide the search. In that sense, it could be related to the transduction done in support vector machines (cf. [Joachims, ICML-99], for instance). Instance based function learning still relies on a distance, but it learns complex outputs (first-order terms) instead of a simple output (such as the value of the class value). Roughly, it considers inputs and outputs represented by first-order, a distance is used to find the examples whose inputs are the nearest-neighbours of the input of the target, then the output of the target is built recursively by a kind of analogical reasoning: the most common functor is chosen among those of the neighbours.

ILP-99 had a new late breaking papers' session. It gave the opportunity to hear from twelve on-going works. Each one was presented in eight minutes plus two minutes for questions. I appreciated this quick overview of on-going work. It allowed me to get a view on even more recent stuff than the regular papers. It also enabled more authors to present their works and get some feedback. I agree that the discussion was short but it was easy to go and talk with the authors. I liked this idea, but some participants complained that it was too short to get a good idea of the work, it was requiring too much concentration, and they suggested that a poster session would have allowed them to focus on some works only. Unfortunately, it wasn't possible since there was already a joint poster session with ICML for the regular papers. I think a poster session -if it is possible- with a five minutes preview of each poster in a plain session at the beginning would be a nice trade-off, even if it is indeed a hard work for the authors, and it takes some time off the session.

As I mentioned earlier, there was a joint poster session with ICML-99, where all regular papers from both conferences were presented. ILP-99's and ICML-99's authors were asked to stand by their posters randomly one out of the two hours allocated to the poster session. Unfortunately each author could only view half of ICML-99's posters and half of ILP-99's posters. It was therefore suggested that, next time, one hour would be devoted to ICML-99 and one hour to ILP-99, since the idea is that both communities look at each other's works. I would like to point out that it was the only flaw, and I would like to take here the opportunity to congratulate the program chairs and the local chairs for the very good organisation. Overall, the joint poster session was very successful.

The proceedings, edited by Saso Dzeroski and Peter Flach, are published in the Lecture Notes in Artificial Intelligence, Number 1634, Springer. Additional material will be put on line at: