Erythemato Squamous Disease Prediction using Ensemble Multi-Feature Selection Approach Efosa Charles Igodan Department of Computer Science, University of Benin, Benin City, Nigeria e-mail charles.igodan@unibe.edu Aderonke Favour-Bethy Thompson Department of Computer Science, Federal University of Technology, Akure, Nigeria e-mail afthompson@futa.edu.ng Olumide Obe Department of Computer Science, Federal University of Technology, Akure, Nigeria e-mail ooobe@futa.edu.ng Otasowie Owolafe Department of Computer Science, Federal University of Technology, Akure, Nigeria e-mail oowolafe@futa.edu.ng Abstract— Skin disease dataset has six classes characterized with redundant and noisy features making classification very difficult due to similarities among the classes. To obtain relevant features to the target concept has being a major component in data mining processes. Although, there is no “one-fit-all” model that outperform all others, in this paper, we described a new methodology that combines four different filtering and three embedded feature selection methods to obtain optimal features for individual models and their ensemble for skin disease. On Dermatology datasets, the proposed ensemble method, which is based on machine learning, was able to classify skin disease types into six categories using the one vs many classification approach. The results show that the stacked ensemble obtained 92.9% accuracy, 85.8% sensitivity and 97.4% specificity compared to both single and ensemble classifiers. This paper proves that ensemble learning methods predict skin disease more accurately and effectively. Keywords: Filtering Method; Embedded Method; Ensemble Learning; SVMs and Erythemato Squamous Disease I. INTRODUCTION The human body is covered with the skin totaling 20 square feet area. The skin protests the body from cold, heat, diseases and regulates the body temperature when and where necessary. The skin can be affected by a disease known as erythemato- squamous either through external and genetic factors. Erythemato-squamous disease can be classified into six classes C1: psoriasis, C2: Soborrheic dermatitis, C3: lichen plarus, C4: pityriasis, C5: chronic dermatitis and C6: pityriasis rubra. It is a difficult task identifying and diagnosing skin disease due to the fact that all classes have similar clinical properties with small changes [17, 43 and 42]. Whether it is fungal, allergic-type disease or the viral type, it’s critical to identify a skin disease in its early stages. [5] so as to prevent and stop further spread to other parts of the skin. To obtain skin disease parameters for examination by dermatologists, 12 clinical attributes are first examined, if symptom of the disease is found then 22 histopathological attributes are also examined by microscopic analysis. The basic treatment and diagnosis of the skin disease is biopsy. The main challenge dermatologists face when diagnosing skin diseases is that the symptoms of one class of disease can overlap with those of other classes, making treatment and diagnosis extremely difficult [43]. With the advent of artificial intelligence, data mining, machine learning algorithms and ensemble learning techniques being introduced into the field of medicine, researchers have used different combinations of these methods for the diagnosis, prediction and classification of various medical diseases thereby improving classification performances [39]. These machine learning algorithms have been applied in breast cancer, lung cancer, liver cancer, kidney disease, Erythemato-squamous disease and many more. Machine learning algorithms help to assist the medical professionals in complex diagnosis, prognosis and patient medical monitoring and planning schemes. However, these algorithms do not perform well if the feature set's size is greater than the instances' size. One of the challenges in data quality is high latitude data containing unimportant features which include redundant, irrelevant and noisy features. These unimportant features negatively can affect the performance of the prediction model and processing time complexity [15] because instances with many irrelevant but noisy features provide very little information [35], Motivated by these limitations, there is the need therefore to obtain optimal performance by applying feature selection method on the medical dataset before classification algorithms are performed. Ensemble method and feature selection are two current hot topics in machine learning that are widely used to improve single learning machine generalization performance [22]. The concept is based on the idea that combining the output of multiple experts is better than a single expert's output [7], resulting in increased diversity and accuracy of the individual base model. Feature selection was used in [6, 7] as a method for generating diversity in classification ensembles. In this case, diversity was incorporated as a goal in the quest for the best collection of feature subsets. Two different criteria are used to classify feature selection techniques. To begin, it is classified as supervised, unsupervised, or semi-supervised based on the prior knowledge available. Second, it can be divided into filter, wrapper, embedded method (54, 51, 14), hybrid, and ensemble methods [17] depending on how they combine the selection algorithm with model building. Therefore, in this work, we present an ensemble-based multi-filter-multi-embedded feature selection (EMFMEFS) method that combines the output of multi-filter methods: Information gain (IG), Gain ratio (GR), Chi-Squared and ReliefF and multi-embedded methods: Recursive Feature Elimination for Support Vector Machine (RFE-SVM) [11], Prediction Risk-based Feature sElection for Bagging (PRIFEB) of SVMs and Mutual Information based Feature sElection for Bagging (MIFEB) to select only descriptive and informative International Journal of Computer Science and Information Security (IJCSIS), Vol. 20, No. 2, February 2022 95 https://sites.google.com/site/ijcsis/ ISSN 1947-5500