Summary

Project MARFIL (Mining data with more flexible representations) has as objective to develop novel approaches for knowledge extraction in those contexts demanding some additional flexibility in data representation:

• Multi-instance and relational learning models that enable a more flexible representation of the input space.
• Learning models with multiple outputs, especially multi-label learning, that allow representing the output space with more flexibility.
• Multi-source and multi-view learning models, which make possible to combine together several data sets describing the same problem using models individually chosen for each of these data sources.

Having all the approaches aforementioned, we will develop new models in the scope of classification, clustering, association and subgroup discovery. We will also enable mechanisms to adapt these models to problems with special characteristics, such as a large number of variables, or very large data sets, as the circumstance dictates. Some of these problems fit into the so-called big data term, and therefore our proposals will be adapted to this new landscape, supplying scalable implementations that are able to provide innovative, appropriate solutions in these contexts.

In addition to its theoretical dimension, previously introduced, this project has got an applied orientation, since we expect to solve several real life problems making use of the developed models. More specifically, we will address some issues related to the context of educational data mining (predicting students' academic performance, modelling self-assessment and peer assessment plans, and developing resource and activities recommendation models for students), and biomedicine (early diagnosis by studying electronic health records, and predicting the risk of insulin metabolism diseases and related pathologies). It is remarkable the interest that nowadays arouses both application fields in our society, as well as the significant impact that any small step forward would have on the health and educational communities. In fact, in addition to our close cooperation with the Universities involved in this project and the Maimónides health research institute, several companies in both sectors have already shown their interest in the results derived from this proposal. Therefore, in a first stage, we will analyse whether these representation models really represent an important step forward to serve the problem resolution with respect to traditional approaches. In a second stage, the existing state of the art methods will be compared to our own proposals, where we expect to achieve significantly improved outcomes.

Last but not least, in order to promote the conducted research, we plan to build test data repositories together with each one of the resulting models in order to allow the scientific community to replicate our experimentation and thoroughly compare the results. Furthermore, we will integrate the developed models into the today's most relevant software platforms in order to facilitate their dissemination.

Research results