nTOPSIS: Developing a multivariate multiclass unsupervised classifier using an ensemble of TOPSIS modules

Accurate image and object classification in precision agriculture research, as well as other fields, has been achieved using supervised and semi-supervised learning methods. These methods require preparing training datasets, training classification models, and tuning their hyperparameters. The methodological contribution of this research is the development of a novel trainingless and hyperparameter-free classification method based on the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS). This research transforms TOPSIS from multi-criteria decision-making (alternative ranking) to unsupervised classification. The novel method, denoted nTOPSIS, uses a network of TOPSIS modules to predict object classes; hence the (n) prefix. The nTOPSIS classifier is developed for offline implementations on edge-computers to solve precision agriculture problems. It serves as a general-purpose, easy-to-use, and lightweight classifier for tabular datasets. The performance of the nTOPSIS classifier is compared to four supervised classifiers: KNN, SVM, RF, and XGBoost, in addition to the zero-shot transformer-based TabPFN. The classification results of four UCI datasets: Iris (and Gaussian-augmented Iris), Palmer Penguins, Rice Cultivars, and Mammographic Mass, are reported. The performance of the nTOPSIS unsupervised classifier was on par with the benchmarking classifiers, achieving weighted accuracy scores of 0.96, 0.98, 0.98, 0.92, 0.81 across the respective datasets without optimization. Also, the nTOPSIS classifier was faster than TabPFN and XGBoost. This work constitutes a novel contribution to unsupervised classification methods. • nTOPSIS is a general-purpose multivariate and multiclass unsupervised classifier. • Combines multiple TOPSIS modules into a classification network. • Provides trainingless, data-independent, and hyperparameter-free classification. • Conforms to the Confusion Matrix performance metrics. • Classifies observations based on the similarity of their attributes to actual class attributes. • Generates class-wise confidence scores akin to supervised classifiers.