Special Issue on Supervised and Unsupervised Classification Algorithms—Foreword from Guest Editors

1. Introduction

Supervised and unsupervised classification algorithms are the two main branches of machine learning. Supervised classification refers to training a system using labeled data being divided into classes, and assigning data to these existing classes. The process consists of computing a model from a set of labeled training data and then applying the model to predict the class label for incoming unlabeled data. This is called supervised learning because the labels are provided by a supervisor, so they are assumed to be correct. Regression is a generalization of classification where the class label is a continuous variable.

In unsupervised classification, the data are unlabeled; therefore, with a lack of prior knowledge, the algorithm tries to search for similarities between points in the dataset. Unsupervised classification algorithms include, but are not limited to, clustering, data estimation, outlier detection, and dimensionality reductions.

Applications range from object detection from biomedical images and disease prediction to natural language understanding and generation.

The Special Issue aims to collect several high-quality research papers on how supervised and unsupervised classification algorithms were modified to overcome the limits of the standard ones, or were explicitly designed to efficiently solve real-life problems. Here, we present some details of the papers’ contents, and we suggest that the reader select interesting articles and explore their findings.

In [1], the authors present KosaNet, a multinomial classification algorithm based on decision trees for analyzing dissolved gases in oil-impregnated power transformers. The transformers are part of the electrical network equipment, and are expensive and responsible for the overall providers’ network reliability and operability. Tests performed on real data provided by the utility power company Kenia Power Ltd show that KosaNet outperforms the state-of-the-art classifiers, especially when dealing with multinomial data.

In [2], a novel temporal symbolic regression method is developed to solve the air-quality modeling and forecasting problems. An actual air-quality database was used for several tests containing measurements from 2015 to 2017 of the Polish city of Wroclaw concerning traffic volume, meteorological parameters, and pollution data. The experimental results show that the proposed strategy has superior statistical performances compared to classical approaches, ranging from simple linear regression methods to recurrent neural networks.

The study in [3] analyzes the effects of nonlinearity on the performance of deep learning methods using the well-known activation functions ReLU and L-ReLU. Several tests were performed on MNIST databases, using different data domains as the network input. The authors empirically showed that ReLU is more effective than L-ReLU when the neural network uses sufficient training parameters. Instead, L-ReLU outperforms the accuracy of ReLU in classification problems. They showed that the information loss due to the activation functions could be evaluated through the entropy function, providing a measure of randomness or disorders in neural network information flow.

A new Fuzzy C-Means (FCM) algorithm is proposed in [4], using multiple fuzzification coefficients to enhance the performance of the basic formulation. The fuzzification coefficient is the exponential parameter of the fuzzy membership functions in the element-cluster squared distances of the objective function to be minimized. Usually, there are no criteria for selecting the coefficient, and it is chosen to be the same for all elements after several tests. Here, the key idea is to evaluate the coefficient based on the surrounding elements’ concentration. High concentration means a high probability of creating a cluster: therefore, a small fuzzification coefficient is selected for faster convergence. Instead, a large coefficient will be chosen for low concentrations to increase the probability of a cluster selection. The experimental tests run on several datasets from UCR showed the new algorithm formulation’s advantages in accuracy and efficiency.

In [5], a new clustering classification algorithm based on dynamic time warping (DTW) was designed to analyze car park data with natural time series characteristics and periodicity. A time series clustering framework comprises four main steps: data preprocessing, distance measurement, clustering, and evaluation. Here, the authors proposed using the weekly periodicity to fill in the missing data values. Secondly, the DTW distance was employed as the distance measurement for the time series data of the car parks. Thirdly, the density-based partition around medoids (DBPAM) clustering method was adopted to the cluster data. Finally, the Purity metric was applied to evaluate the results. Several tests were conducted using 9 UCR datasets and data collected from 27 car parks operated by the Birmingham City Council NCP during two months of 2016. The results showed that the designed classifier exhibited a better performance than those based on Euclidean distance measurement and traditional clustering models based on DTW.