Improving the Accuracy of Features Weighted k-Nearest Neighbor using Distance Weight

Abstract

: FWk-NN is an improvement of k-NN, where FWk-NN gives weight to each data feature thereby reducing the influence of features that are less relevant to the target. Feature weighting is proven to be able to improve the accuracy of k-NN. However, the FWK-NN still uses the majority vote system for class determination to new data. Whereby the majority vote system is considered to have several weaknesses, it ignores the similarity between data and the possibility of a double majority class. To overcome the issue of vote majority at FWk-NN, the research will change the voting majority by using distance weight. This study uses a dataset obtained from the UCI repository and a water quality data set. The data used from the UCI repository are iris, ionosphere, hayes-Roth, and glass. Based on the tests carried out using UCI repository dataset it is proven that FWk-NN using distance weight has averaged an increase about2%, with the highest increase of accuracy of 4.23% in the glass dataset. In water quality data, FWk-NN using distance weight can achieve an accuracy of 92.58% or has increased 2% from FWk-NN. From all the data tested, it is proven that the distance weight is able to increase the accuracy of the FWk-NN with an average increase about 1.9%.