Training Neural Networks Using Input Data Characteristics

• Main Author: CERNAZANU, C.
• Format: Article
• Language: English
• Published: Stefan cel Mare University of Suceava 2008-06-01
• Series: Advances in Electrical and Computer Engineering
• Subjects: neural networks; data mining; correlation-based feature subset selection method; data features extraction; training algorithm
• Online Access: http://dx.doi.org/10.4316/AECE.2008.02012
• ISSN: 1582-7445; 1844-7600

Feature selection is often an essential data processing step prior to applying a learning algorithm. The aim of this paper consists in trying to discover whether removal of irrelevant and redundant information improves the performance of neural network training results. The present study will describe a new method of training the neural networks, namely, training neural networks using input data features. For selecting the features, we used a filtering technique (borrowed from data mining) which consists in selecting the best features from a training set. The technique is made up of two components: a feature evaluation technique and a search algorithm for selecting the best features. When applied as a data preprocessing step for one common neural network training algorithms, the best data results obtained from this network are favorably comparable to a classical neural network training algorithms. Nevertheless, the first one requires less computation.