missing completely at random data
minimum spanning tree
Artificial neural networks with local approximation abilities become more common approach for a variety of tasks due to their tolerance to a high-dimensional data and large training set volumes. However, missing values in a training set is often a problem for a wide known class of such networks – Radial Basis Network (RBF). Common ways for dealing with missing data is often not an option as an effective solution for that problem.
The work describes a new model for RBF neural network parameters adjustment that based on static clustering approach. Apart from being quite good for handling with missing completely at random data, this method provides an alternative solution for the optimal estimation of basis function number. Radial neuron centers can also be effectively assigned. The model is based on both computational matrixes and cluster data recycling, which were set along with the main cycle of data imputation process. Particularly, two additional minimum spanning tree-based clustering procedures were applied for both large clusters partition, and small clusters merge. This is due to fix some problems with suboptimal setting for clustering algorithm parameter, and processing data characteristics.
The higher overall efficiency of the proposed model compared with traditional RBF neural network was confirmed by a majority of the experimental results on public domain datasets with different fraction of missing data.