Feature ranking based on an improved granular neural network

In this paper, we try to solve the feature ranking problem through an allocation of information granularity. In many real applications, people are more concerned with an ordered sequence, especially a sequence with a few most important features. However, the outcome of the feature selection methods is often not stable. We proposed an improved granular neural network framework to provide a comparable stable ordered sequence. Unlike other granular neural networks, this network uses information granules as input and generates granular output which is optimized with higher generality and specificity. This way, the construction of the sequence of ordered features is realized from a more comprehensive perspective (neither regression nor classification). The information granules are formed by allocating a level of information granularity onto numeric features and then being optimized through an optimization tool (genetic algorithm for instance). Computational experiments on both synthetic and real data sets are performed to compare the stability of our algorithm. The results show consistency with experts’ suggestions.