Abstract. This paper investigates a general framework for learning concepts that allows to generate accurate and comprehensible concept representations. It is known that biases used in learning algorithms directly affect their performance as well as their comprehensibility. A critical problem is that, most of the time, the most "comprehensible" representations are not the best performer in terms of classification! In this paper, we argue that concept learning systems should employ Multiple-Knowledge Representation: a deep knowledge level optimised from recognition (classification task) and a shallow one optimised for comprehensibility (description task). Such a model of concept learning assumes that the system can use an interpretation function of the deep knowledge level to build an approximately correct comprehensible description of it. This approach is illustrated through our GEM system which learns concepts in a numerical attribute space using a Neural Network representation as the deep knowledge level and symbolic rules as the shallow level.

No abstract

This article presents VADIS Consulting’s solution for the cross-selling problem of the PAKDD_2007 competition. For this competition, we have used our in-house developed tool RANK, which automates a lot of important tasks that must be done to provide a good solution for predictive modeling projects. It was for us a way of benchmarking our 3 years of investment effort against other tools and techniques. RANK encodes some important steps of the CRISP-DM methodology: Data Quality Audit, Data Transformation, Modeling, and Evaluation. We have used RANK as we would do in a normal project, however with much less access to the business information, and hence the task was quite elementary: we have audited the data quality and found some problems that were further corrected. We have then let RANK build a model by applying its standard recoding, and then applied automatic statistical evaluation for variable selection and pruning. The result was not extremely good in terms of prediction, but the model was extremely stable, which is what we were looking for.

This paper presents a general learning structure called ILISCE which learns meta-rules to guide a scheduling system. The scheduling program, called OPAL, solves real-world problems using a set of local heuristics which incrementaly resolve the whole set of conflicts among the physical resources of the floor. A Selective Inductive learning approach is used to create concepts representing typical states of the job shop which allow a Classifier to "recognize" current situations and to choose the corresponding best operators to apply on. In this paper, we present the general architecture of ILISCE.

No abstract