Classification, game theory and feature engineering: new models and applications

About

The classification problem is a central one in machine learning with multiple applications in economics, finance, politics, medicine, engineering, biology, etc. During the last years more emphasis has been placed on large data rather than focusing on details, leading to possible loss of important information. Recent increase in computational power allows the exploration of new solution concepts, providing a more nuanced insight into the classification. Game theory offers a plethora of equilibria concepts designed for various strategic situations for which the concept of optimality does not work. The main goal of this project is to explore and improve the use of game theoretic concepts in solving supervised and unsupervised classification problems with large number of features. The supervised classification problem is converted into a game, in which data instances are players that choose their class such that the Nash equilibrium of the game represents the correct classification. Parameters of probabilistic models can be estimated to approach the equilibrium of the game. The model can be evolved by using from mechanism design and genetic programming. Genetic programming tools will also be used for automated feature engineering. The unsupervised classification problem will be approached by converting the data into networks and identify clusters/communities and important features/critical nodes by combining computational intelligence tools with suitable equilibria concepts.