Functional Expansions Based Multilayer Perceptron Neural Network for Classification Task

Abstract

Artificial neural network has been proved among the best tools in data mining for classification tasks. Where, Multilayer Perceptron (MLP) is known as benchmarked technique for classification tasks due to common use and easy implementation. Meanwhile, it is fail to make high combination of inputs from lower feature space to higher feature space. In this paper, Shifted Genocchi polynomials and Chebyshev Wavelets functional expansions based Multilayer Perceptron techniques with Levenberg Marquardt back propagation learning are proposed to deal with high dimension problems in classification tasks. Five datasets from UCI repository and KEEL datasets were collected to evaluate the performance in terms of five evaluation measures. Ttest was applied to check the significance of the proposed techniques. The comparison results show that the proposed models outperform in terms of accuracy, sensitivity, specificity, precision and f-measure.

A Classification Methodology Based on Subspace Graphs Learning

“…Finding the best partitions is a hard task but necessary to finalize the classification objective. In this work, the methodology proposed has the aim to find out best TABLE IV: Comparison in terms of sensitivity, specificity, precision and accuracy between our model (SPT CD) and results of three variations of Multilayer Perceptron in [28]. partitions with ensemble method from the neighbourhood of an instance, based on subspace graph in training step and use this information as parameter in the objective function to improve final accuracy.…”

“…Comparison in terms of sensitivity, specificity, precision and accuracy between our model (SPT CD) and results of three variations of Multilayer Perceptron in[28]. Acronyms in this Table are defined in I.…”

“…12: end for 13: Final majority vote (C ≥ 0) and assign prediction Algorithm 4 SPT CD (h, h', z) The approach is equal to lines [20][21][22][23][24][25][26][27][28][29][30][31] 34: end if 35: return Furthermore, in Fig 1 we consider a boundary decision derived from the weighted edge sum of a specific spanning trees (blue and green boundaries). The reason we don't consider only the edge couples among k-neighbours of instance nearest to it (in training and test step) is to create a boundary more detailed thanks to his neighbourhood.…”