Supervised learning vector quantization for projecting missing weights of hierarchical neural networks

Abstract

A supervised learning vector quantization (LVQ) method is proposed in this paper to project stratified random samples to infer hierarchical neural networks. Comparing with two traditional methods, i.e., list-wise deletion (LWD), and non-amplified (NA), the supervised LVQ shows satisfying efficiencies and accuracies in simulation studies. The accomplishments of proposed LVQ method can be significant for sociological and psychological surveys in properly inferring the targeted populations with hierarchical neural network structure. In the numerical simulation study, successes of LVQ in projecting samples to infer the original population are further examined by experimental factors of sampling sizes, missing rates, and disproportion rates. The experimental design is to reflect practical research and under these conditions it shows the neural network approach is more accurate and reliable than its competitors.