The 20 th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning
ESANN 2012
Bruges, Belgium, 25-27 April 2012

Special session: Interpretable Models in Machine Learning

 
Organizers:

Paulo Lisboa, Department of Mathematics & Statistics, Liverpool John Moores University, Liverpool L3 3AF, U.K., p.j.lisboa@ljmu.ac.uk
Alfredo Vellido, Department of Computing Languages and Systems, Technical University of Catalonia, Barcelona 08034, Spain. avellido@lsi.upc.edu
José D. Martín, Electronic Engineering Department, University of Valencia, Valencia 46100, Spain. jose.d.martin@uv.es

 
Description:

Practical application of machine learning algorithms for knowledge extraction requires interpretability of the analytical models, without which it is difficult to validate against domain expertise and to explain the extracted knowledge to the user. General linear models can be expressed in terms of the relative contribution of each individual data attribute, but it is not obvious how to extend this concept to non-linear models. This problem of interpretability extends to all machine learning fields (classification, prediction, clustering, etc.) and could severely limit the general adoption of non-linear models.

On the other hand, multivariate association mining with graphs and networks provides alternative avenues towards interpretation of complex data.

Interpretability can also be enhanced through, for instance, data visualization and rule extraction, among other approaches. This session is focused on basic methodology for the interpretation of efficient non-linear models including, but not limited to, the following:
i. interpretation of non-linear models, including SVMs and other kernel methods
ii. inductive learning, including rule generation from data and interpretation of random forests
iii. structure finding, including efficient derivation of directed graphs
iv. deep learning
v. nonlinear dimensionality reduction and prototype based learning, including data visualization
vi. causal models, including integration of prior knowledge into data-based models

Theoretical models should be illustrated, whenever possible, by real-world examples, cross-checking the interpretation of the model using prior knowledge or demonstrating the discovery of new insights about the problem domain. The ultimate objective is to obtain generic non-linear models that may be complex to build but can gain acceptance by non-expert users.

Dates:
Submission of papers:  30 November 2011
Notification of acceptance: 23 January 2012