An Ar2p Deep Learning Architecture for the Discovery and the Selection of Features

Abstract

In the context of pattern recognition processes with machine learning algorithms, either through supervised, semi-supervised or unsupervised methods, one of the most important elements to consider are the features that are used to represent the phenomenon to be studied. In this sense, this paper proposes a deep learning architecture for Ar2p, which is based on supervised and unsupervised mechanisms for the discovery and the selection of features for classification problems (called Ar2p-DL). Ar2p is an algorithm of pattern recognition based on the systematic functioning of the human brain. Ar2p-DL is composed of three phases: the first phase, called feature analysis, is supported by two feature-engineering approaches to discover or select atomic features/descriptors. The feature engineering approach used for the discovery, is based on a classical clustering technique, K-means; and the approach used for the selection, is based on a classification technique, Random Forest. The second phase, called aggregation, creates a feature hierarchy (merge of descriptors) from the atomic features/descriptors (it uses as aggregation strategy the DBSCAN algorithm). Finally, the classification phase carries out the classification of the inputs based on the feature hierarchy, using the classical Ar2p algorithm. The last phase of Ar2p-DL uses a supervised learning approach, while the first phases combine supervised and unsupervised learning approaches. To analyze the performance of Ar2p-DL, several tests have been made using different benchmarks (datasets) from the UCI Machine Learning Repository, in order to compare Ar2p-DL with other classification methods.