Machine learning with UMAP k-NN-ML classification for data processing: a case study on LocURa4Iot dataset

Abstract

The interest of the classification problem in IoT networks has recently been highlighted by numerous researchers. Various classification and learning algorithms are used to determine the precise locations of nodes. This paper introduces a supervised k-nearest neighbor (k-NN) classification model to categorize new nodes in Internet of Things collection networks. The proposed method uses real training datasets and shows that the k-NN classifier can reduce misclassification costs efficiently. The LocURRa4Iot dataset, with two classes of data, is used. The first class, called class1, contains two attributes. The first attribute represents the strength indicator of the signal received from the first message (rssiRequest), and the second attribute represents the signal strength indicator received from the second message (rssiAck). The second class, called class2, contains two attributes: flight time and flight time after correcting the clock offset on the second message (ACK). These classes are used as training data. The aim is to assign new nodes to the training classes using the k-NN algorithm. The obtained results indicate that the UMAP k-NN-ML algorithm. The obtained results indicate that the proposed classifier with \(\varvec{k}\varvec{=}\varvec{3}\varvec{,} \varvec{5}\varvec{,}\) and \(\varvec{7}\) reveals a trade-off between local sensitivity and noise robustness. The \(\varvec{k}\varvec{=}\varvec{5}\) setting delivers the best performance, balancing precision (87%), recall (85.8%), and F1-score (86.8%). While \(\varvec{k}\varvec{=}\varvec{3}\) is more sensitive to fine patterns but prone to noise, \(\varvec{k}\varvec{=}\varvec{7}\) tends to underfit due to excessive smoothing. Overall, \(\varvec{k}\varvec{=}\varvec{5}\) offers the optimal compromise for this dataset.