International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056 Volume: 04 Issue: 09 | Sep -2017 www.irjet.net p-ISSN: 2395-0072 © 2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1385 Enhancing Classification Accuracy of K-Nearest Neighbours Algorithm Using Gain Ratio Aditya Duneja 1 , Thendral Puyalnithi 2 1 School Of Computer Science and Engineering , VIT University, Vellore, 632014 , Tamil Nadu, India 2 Professor, School Of Computer Science and Engineering, VIT University, Vellore, 632014, Tamil Nadu, India ---------------------------------------------------------------------***--------------------------------------------------------------------- Abstract – K-Nearest Neighbors or KNN classification technique in machine learning is among the simplest of techniques which are used for classification. The approach used in KNN involves finding the k nearest neighbors to a test vector and then assigning a class to it based on majority voting. One of the assumptions in KNN is that data lies in a feature space, that is, measures like Euclidean distance, Manhattan distance and others can be applied on data. It has its applications in computer vision, recommender systems and various other fields. In this paper, a modification to this method has been put forward which aims at improving its accuracy. The modified version combines the concepts of distance metrics, which are applied in K-nearest neighbors, and the entropy or the strength of an attribute. The way in which the distance is calculated between the data points is altered as the strength of each individual attribute or dimension is taken into consideration. The proposed technique was tested on various UCI datasets and the accuracies of both methods, the old and the modified, are compared. Key Words: classification, entropy, Euclidean distance, KNN, dimension 1. Introduction In machine learning, classification is a method which uses a training set comprising of various instances and their respective classes to determine the category of a new observation or a set of new observations. It is a supervised learning technique. Classification has a wide range of applications in various fields, namely medical imaging, speech recognition, handwriting recognition, search engines and many more. So, high accuracy classification models are required in order to make correct predictions in the above mentioned fields. Some of the classification models are decision trees, Support Vector Machines and K- nearest neighbor classifier. K-nearest neighbor or KNN classifier is one of the most frequently used classification techniques in machine learning. K nearest neighbor algorithm involves the computation of the distance of the test observation with every observation in the training set. The distance which is computed is mostly Euclidean distance although other measures of distance can also be applied. Then the k nearest observations are chosen and the class which is assigned to the majority out of these k observations is assigned to the test observation. KNN makes no assumptions regarding the distribution of data, therefore, it is non-parametric. No separate training phase is defined for this algorithm as the whole dataset is used for computation of distance every time. KNN has undergone many modifications over the years in order to improve its accuracy. Faruk et al.[1] proposed new version of KNN where the neighbors are chosen with respect to the angle between them. Even if two neighbors have the same distance from the test vector they are considered different with respect to the angle they make. Singh et al. [2] devised a way to accelerate the calibration process of KNN using parallel computing. Parvin et al. [3] modified the KNN algorithm by weighting the neighbors of the unknown observation. The contribution of every neighbor is considered to be inversely proportional to its distance from the test vector. Fuzzy K nearest neighbors was first proposed in 1985 by Keller et al.[4] Fuzzy memberships are assigned to the training data in order to obtain more accurate results as every neighbor is given a different weight. A new modification to the K-nearest neighbors algorithm is proposed in this paper taking the strength of the attributes into account. Faziludeen et al.[5] proposed a new modification to KNN known as the Evidential K nearest neighbors (EKNN). The resultant class from all k nearest neighbors of a test sample were used to find the class, not only the majority class. Distance between samples was considered to measure the weight of the evidence attributed to each class. Some research is also done to reduce the size of the dataset required to get effective results from the KNN algorithm. Song et al.[6] proposed an algorithm which removed the outlier instances from the dataset and the remaining were sorted on the basis of difference in output among instances and their neighbors. Another modification to KNN was put forward by Nguyen et al.[8] in the form of a distance metric learning method. The resulting problem was formulated as a quadratic program. Lin et al.[9] proposed a nearest neighbor classifier by combining neighborhood information. While finding the distance between samples, the influence of their neighbors is taken into account. Manocha et al. [10] proposed a probabilistic nearest neighbor classifier which used probabilistic methods to assign class membership, an aspect which KNN doesn’t have. Sarkar[11] improved the accuracy of the KNN by