Graph-Based Weighted KNN for Enhanced Classification Accuracy in Machine Learning

Abstract

The K-Nearest Neighbor (KNN) algorithm is widely recognized for its simplicity and effectiveness in classification tasks. However, it is notably sensitive to outliers and requires careful tuning of the parameter k. To address these limitations, we propose a Graph-based Weighted KNN (GBWKNN) algorithm that leverages the structural properties of KNN graphs to assign dynamic weights to neighbors based on their mutual connectivity and relevance. Unlike conventional KNN, which relies solely on distance metrics, our model incorporates inward connection strength to mitigate the influence of noise and outliers. Through experimentation using the UCI Wine dataset, our proposed approach demonstrated consistently improved classification accuracy over standard KNN across various k values. Additionally, the model integrates a weight normalization mechanism, enhancing the robustness of predictions. Comparative analysis indicates an average performance improvement of up to 5% in classification tasks. This research signifies a meaningful step towards adaptive learning mechanisms within traditional classifiers and opens new avenues for robust, graph-enhanced machine learning models.