Affiliations: [a] Department of IT, Shaheed Benazir Bhutto University, Shaheed Benazirabad, Sindh, Pakistan
| [b] Deparment of Computer and Information Sciences, Universiti Teknologi PETRONAS, Malaysia
| [c] The College of Computer Science and Engineering in Yanbu, Taibah University, Medina, Saudi Arabia
| [d] Informatics Research Institute, The City for Scientific Research and Technology Applications, Alexandria, Egypt
| [e] Department of Computer Science, SMIU, Karachi, Pakistan
| [f] Department of Computer Systems Engineering, Faculty of Engineering, The Islamia University Bahawalpur, Pakistan
| [g] Department of Information Systems and Technology, Faculty of Computer Science and Engineering, University of Jeddah, Jeddah, Saudi Arabia
Correspondence:
[*]
Corresponding author. Munwar Ali, Department of IT, Shaheed Benazir Bhutto University, Shaheed Benazirabad, Sindh, Pakistan. E-mail: mazardari@gmail.com
Abstract: The k-NN algorithm is an instance-based learning algorithm which is widely used in the data mining applications. The core engine of the k-NN algorithm is the distance/similarity function. The performance of the k-NN algorithm varies with the selection of distance function. The traditional distance/similarity functions in k-NN do not perfectly handle the mix-mode words such as when one string has multiple substrings/words. For example, a two-word string of “Employee Name”, a one-word string of “Name” or more than one word such as, “Name of Employee”. This ambiguity is faced by different distance/similarity functions causing difficulties in finding the perfect match of words. To improve the perfect-match calculation functionality in the traditional k-NN algorithm, a new similarity distance metric is developed and named as word-distance (w-distance). The perfect match will help us to identify the exact required value. The proposed w-distance is a hybrid of distance and similarity in nature because it is to handle dissimilarity and similarity features of strings at the same time. The simulation results showed that w-distance has a better impact on the performance of the k-NN algorithm as compared to the Euclidean distance and the cosine similarity.
Keywords: k-NN algorithm, distance/similarity metric, text match, data mining, cosine similarity
