Research in machine learning and data privacy

This is an application for a research programme in the area of Machine Learning, with the focus on two specific topics, i.e. data privacy and learning from text data. It is clear that sharing of the distributed, web-based data, as well as cloud computing, make data privacy one of the main challenges for Computer Science. Privacy-preserving Data Mining (PPDM) is an active research area, developing technical solutions to these data privacy challenges, particularly in a data mining context. I have been active in PPDM for the last ten years. I propose here to develop several new PPDM techniques. In one approach, the data is modified in a distributed manner, so that private data is hidden and cannot be easily linked back to an individual. In another approach, the data is masked using state of the art cryptographic tools ("zero-loss privacy"). I also propose to work on a definition of data privacy which will encompass data aspects left out of the existing definitions. In particular, I want to integrate the cost of the attack and the cost of the protection into a definition of privacy. The second part of the proposed program is devoted to learning from text data. The major focus is on a light, linguistically-informed data model beyond the existing bag of words (BOW) representation. I want to break away from the constraints of the BOW model and work on learning with a more expressive, and yet realistic First Order Logic-based representation. On the one hand, relational learning from logical representations does not scale up as well as learning from the standard propositional (attribute-value) representations of the data. On the other hand, recent progress of Inductive Logic Programming (my former area of research) in relational learning makes me believe that it will soon be feasible to perform large-scale inductive learning from such representations of text data. I propose a research program leading towards that goal.