Using weight constraints and masking to improve fuzzy cognitive map models

Abstract

The paper presents a novel supervised learning method for fuzzy cognitive maps adapted from the theory of artificial neural networks. The main objective in designing the method was to pay closer attention to the distinctions that exist between fuzzy cognitive maps, and the original model for which the method was intended – whether it was a feedforward neural network, a recurrent network, or an energy-based model. The augmented version strives to properly build upon the various strengths of fuzzy cognitive maps – particularly on their interpretability, which arises from the close coupling that exists between their nodes and particular concepts. It is shown that the augmented method is able to outperform existing approaches. Notably, the ability of the learned model to generalize correctly, and to faithfully reconstruct the original system is studied.