Evolution and performance of classifiers based on Fuzzy cognitive maps

Abstract:

Concepts and causal relations. FCMs allow decision-makers to analyse the interdependencies and causal relationships between different variables in a system, and also allow researchers to make classifiers. FCMs are useful for building more interpretable classifiers than most state-of-the-art models. But it is also a fact that the interpretability of FCMs causes, as a counterpart, relatively higher classification errors than other models.  In this keynote recent improvements are shown and explained. Firstly, it has been demonstrated that FCMS are not universal approximators. Therefore, the accuracy and approximation capabilities of FCMs are substantially reduced. We present a new learning methodology for FCM-based classifiers that uses the estimated boundaries for neurons. Secondly, for supporting the interpretability without affecting the performance and for avoiding convergence we propose a new reasoning rule that introduces a nonlinearity coefficient φ controlling the extent to which the model will consider the value produced by the transfer function over the neuron’s initial activation value. Thirdly, a new learning method is used based on backpropagation in conjunction with a new loss function. This loss function incorporates the model’s convergence and the errors made. Fourthly, we propose a long-term cognitive network (LTCN) for interpretable pattern classification of structured data. Numerical results show that our new FCM based classifiers outperforms the classical model for all datasets used in the experiments. The work presented has been a collaboration with researchers from Hasselt University (Belgium), Tilburg university (Netherlands) and Universidad Central “Marta Abreu” de Las Villas (Cuba).