27th International Conference on Inductive Logic Programming
4-6 Sep 2017 Orléans (France)
Marc Boullé : Automatic Feature Construction for Supervised Classification from Large Scale Multi-Relational
Data In large telecommunication companies like Orange, data are collected at a peta-byte scale, with a variety of domains ranging from network design, ergonomy, text and web mining to customer relationship management. Given the vast need for many data mining tasks, the issue is to propose methodologies and tools to industrialize as far as possible the data mining process. The raw data most often come with a relational structure, for example with customers in a main table and their call detail records (CDR) in a secondary table. This kind of data requires a heavy phase of data preparation, involving feature selection and construction. In Orange Labs, we have developed an approach to automate feature construction in the multi-relational data mining setting. In this setting, domain knowledge is specified by describing the structure of data by the means of attributes, tables and links across tables, and choosing construction rules. Mining relational data implies to be able to learn complex features aggregating properties of related objects. The space of features that can be constructed is virtually infinite, which raises both combinatorial and over-fitting problems. A prior distribution is introduced over all the constructed features, as well as an effective algorithm to draw samples of constructed features from this distribution. Extensive experiments show that the approach is robust and efficient, outperforms the state of the art and can deal with a nowadays large scale industrial problem. This approach is available in a tool named Khiops, widely used in Orange for mining large scale multi-relational databases. Data mining studies can now be completed in hours, not weeks.
Alan Bundy : Can Computers Change their Minds?
Autonomous agents require models of their environment in order to interpret sensory data and to make plans to achieve their goals, including anticipating the results of the actions of themselves and other agents. These models must change when the environment changes, including their models of other agents, or when their goals change, since successful problem solving depends on choosing the right representation of the problem. We are especially interested in conceptual change, i.e., a change of the language in which the model is expressed. Failures of reasoning can suggest repairs to faulty models. Such failures can, for instance, take the form of inferring something false, failing to infer something true or inference just taking too long. I will illustrate the automated repair of faulty models drawing both on work multi-agent planning and on the evolution of theories of physics.