Classification Performance of Committee Networks Improvement under Sparse Data Conditions

Abstract

In most real world applications, the data for modeling is normally sparse. This makes it difficult for modelers to construct a neural network model. Eventually the training process may cause overfitting. This paper proposes committee network methodology to deal with sparse data for a classification problem. The committees are developed based on bootstrapped training sets and are called adjusted pair-wise committee and adjusted random-mix committee. We test the committees’ performance against that of the bootstrap committee and single neural network using the selected data sets from UCI Machine Learning Repository, Center for Machine Learning and Intelligent System. The results reveal that the proposed models perform as well as or better than the baseline models.