CALL for Papers
Professor Goran Klepac
Head of Chair for Information Systems
University College for Applied Computer Engineering Algebra
Convulsion Neural Networks as A Tool For Early Warning System Development
Paper will propose solution based on convulsion neural networks as a tool for early warning system development. Convulsion neural networks are powerful tool in domain of image recognition, handwriting recognition and other hard learning problems.
Early warning systems are challenging area, especially because of complexity and nature of early warning systems. We cannot observe early warning systems as traditional predictive models, because we are not talking about specific event which we try to predict and simply declare as bad event. There are numerous factors or their combinations in risk driven business which can trigger suspicious about potential future problems for specific elements within portfolio or within part of portfolio. Problem is, because those factors should not be statistically significant on portfolio level, which is a criterion for variable selection if we are talking about classical predictive models.
On the other hand we are talking about different types of riskiness which cannot be unlikely defined under single flag.
One possible solution for this kind of problem could be usage of fuzzy expert systems with which we can vault problems with statistically insignificant factors, which are recognised through expert knowledge. That approach demands leaning on expert knowledge and forward thinking about all potential sources of riskiness.
Taking in consideration convulsion neural networks methodology and their power to concentrate on features, paper will present solution for early warning system in business based on convulsion neural networks.
That means include specific approach to data preparation and attribute selection as well as attribute preparation, with scope on behavioural variable construction an trial and error approach on final solution construction.
Proposed approach combine traditional attribute relevance analysis for recognition of most important variables which can be part of solution, as well as variable selection upon expert judgement and including it into final model.
That approach assure combination and mixing features from statistically significant variables due to preselected observed target variables and from statistically insignificant variables which has great impact from perspective of expert judgement as factors which has high ponder from perspective of riskiness even it is not frequent within data sample.
Paper will shortly present solution for early warning system based on convulsion neural network, which unite traditional predictive modelling with elements of expert approach integrated within convulsion neural networks.
This approach is based on convulsion neural networks which have advantages to focus on characteristic features. Taking in consideration nature of early warning systems which are suitable for solving uncertain and unpredictable real-world problems convulsion neural networks can be good choice for early warning system construction.
Abstract submission: April 4 2018
Notification and Registration opens: May 6 2018
Camera-ready deadline: June 10 2018