This paper focuses on the problem of network and internet service security using a machine learning approach, in the context of class imbalance. The problem of class imbalance is a crucial one to consider when a knowledge discovery problem is targeted towards nuances in the dataset (untypical internet activities), and a machine learning algorithm trained on an imbalanced dataset is biased towards the majority class rather than to the minority class. In such cases, nuances or incidents such as network attacks may not be detected. In this paper, selected oversampling approaches are validated, and we consider an approach in which redundant oversampling and then undersampling are applied. Finally, using benchmark datasets related to cyber-attack problems, selected machine learning models are applied and conclusions are drawn from the results.

Class imbalance remains a pressing issue in machine learning, particularly when advanced technologies require it from machine learning implementations. Medical or cybersecurity applications are good examples of this, as well as others. In this paper, we focus on cybersecurity, where the minority class often represents rare but critical attack instances. We present a comparative study of various oversampling strategies—including SMOTE-based methods, GAN-based methods, and hybrid approaches—for detecting network intrusions in imbalanced datasets. Selected classifiers are evaluated using metrics such as Accuracy, AUC, F1-score (for the minority class), and Recall. The experimental results are analysed and discussed, in terms of the strengths and weaknesses of each oversampling strategy.

The problem of selecting the best convolutional neural network (CNN) architecture appears when dealing with a large number of classes and images. The selection problem resulted in the consideration of varying layer configurations and strategies, and the reliability of the individual deep learning models was unknown. On the other hand, as CNN architectures evolved, each new iteration proposed, from a theoretical point of view, was a more effective method for achieving better results. Thus, this was the motivation to validate the selected CNNs on image recognition and classification. This paper focuses on classifying the International Code of Signals (INTERCO) flags using different CNN architectures, including AlexNet, VGG-16, VGG-19, InceptionV3, ResNet-18, ResNet-34, ResNet-50, MobileNetV2, EfficientNet-B0, EfficientNet-B1, CSPNet, and ConvNeXt-Tiny. The performance validation of these architectures through the analysis of metrics such as the accuracy, precision, recall, F1-score, training time, number of epochs, and single-image processing time has been carried out. The results of the computational experiments are presented and discussed.

Automatic Identification System is a telecommunication system that enables ships to communicate with each other. Exchanged AIS messages contain critical information for early ship collision detection, such as vessel’s position, speed, course, etc. However, due to some technical limitations of AIS (for example, a packet collision phenomenon), some fragments of AIS messages might get damaged during the transmission. In order to maintain the usefulness of AIS data, that corresponds to the maritime safety and security, the reconstruction of damaged AIS data is necessary. This paper discusses a dedicated 3-stage machine learning based framework for AIS data reconstruction. The utilized models learn the relationship between the features in AIS data, which allows for the detection of the abnormal (i.e. damaged) fields of AIS messages and prediction of their correct values. The aim of the paper is to provide detailed insights into the performance of models on the prediction stage, which is the final stage of the proposed framework. The computational experiment, conducted using real AIS data, involved examination of the impact of the position of the damaged bits in a message and the duration of data capturing on the reconstruction quality.

The International Code of Signals (INTERCO) consists of codes and signals used to communicate messages regarding the safety of navigation. These codes are used by vessels, and the signals can be sent in various ways, including flaghoist, signal lamp, flag semaphore, or as a radio message. This paper focuses only on flag signals, proposing their detection and classification using deep learning algorithms. The paper also addresses a specific scenario, where the detection and classification of flag signals are necessary for verifying the current status, operations, and behaviours of vessels, in line with the intention conveyed by the flag signal. The system discussed is regarded as a component of a larger system for ensuring navigation safety and monitoring ship behaviours near critical infrastructure such as ports, offshore wind farms, or other offshore constructions. To tackle this challenge, a convolutional neural network (CNN) has been employed. The paper delves into the process of preparing the image dataset, including annotation and augmentation, and discusses the CNN learning process based on the YOLO algorithm implementation, finally discussing the results of the computational experiment.

Automatic Identification System (AIS) is a telecommunication system that enables ships to communicate with each other. Vessels exchange information about their position, speed, course, etc. Unfortunately, due to some technical limitations of AIS (for example, a packet collision phenomenon), some fragments of AIS data might get damaged during the transmission, which could potentially lead to two ships colliding, not knowing that they are on a collision course. In order to maintain the maritime safety and security, there is a need for the reconstruction of damaged AIS data. This paper discusses a dedicated framework for AIS data reconstruction. Nevertheless, the aim of the paper is to provide detailed insights into the prediction stage, which is one of the three stages of the AIS data reconstruction process. The implementation and validation of two models for predicting values of the damaged fields in vessels’ trajectories are presented. These models include a Vector Autoregression model and an XGBoost model. The experiment consists of two scenarios: evaluating the models’ performance independently and as part of the entire AIS reconstruction framework.

In this paper, the problem of imbalanced data stream classification is considered. To eliminate the problem of imbalanced data, over- and under-sampling procedures are implemented within a dedicated framework called Weighted Ensemble with one-class Classification and Over-sampling and Instance selection (WECOI). To increase the quality of synthetic instance generation (carried out under the umbrella of an over-sampling procedure), an agent-based population algorithm is proposed. The problem is defined, and the agent-based algorithm is presented. A discussion of selected computational results is also included.

This paper describes an application dedicated to collecting and mining reports of software safety vulnerabilities and exposures in physical systems. This work focuses on the clustering problem of such reports, which means grouping them through automated computing process. The clustering is carried out in two stages. In the first stage, potential similarities between the reports, together with the number of clusters are detected through automated text analysis. In the second stage, a hierarchical clustering is conducted to reduce the number of these clusters to provide potential number of appropriate clusters of the group of reports. The clustering of the second stage provides the user with greater flexibility in viewing individual reports. This paper focuses mainly on the first stage of described clustering method. Two selected clustering algorithms have been compared with the aim to show how to detect the most appropriate number of groups between scraped documents. The computational experiment results are presented and discussed in the experimental section of this work.

This paper presents a framework for text clustering and categorisation. The proposed clustering approach is based on a modified existing similarity-based clustering algorithm, which was originally developed for well-structured data. In this study, the clustering algorithm is used to map text documents into clusters, in order to discover groups of topical documents. The clusters produced in this way are also used for the categorisation of new documents that are uploaded to the system. The algorithms are discussed using as an example the analysis of text documents including Industrial Control Systems (ICS) Advisory Reports and Common Vulnerabilities and Exposures (CVE) recommendations, together available and provided by the Cybersecurity and Infrastructure Security Agency (CISA). Experiments are carried out, although the main focus is on the clustering algorithm. Based on the experimental results, it can be concluded that the proposed similarity-based clustering algorithm can be considered as an alternative approach for text clustering.

This paper considers the problem of selecting a state-made electronic document management system (EDMS) using the analytic hierarchy process (AHP) model. The problem of EDMS selection from two available variants is analysed and the selection criteria related are highlighted on the basis of the literature and the authors’ experience. Calculations related to pairwise comparisons and prioritisation between criteria and variants form the core of the paper. This problem is only applicable to municipalities in Poland that have implemented an older state-made system; based on pair-wise comparisons, retaining the older system is preferable to implementing the new state-made EDMS because of the lack of integration of the new EDMS with internal systems.

The paper focuses on the problem of instance selection. Instance selection is currently crucial to enhance the efficacy and efficiency of machine-learning tools when they are used to solve a data-mining task and when the data are large and they are seen through the prism of the big data phenomenon. Instance selection eliminates redundant instances and thus reduces the size of the training data set. The training data, with redundant cases removed, can be more useful and ensure better performance of the final classification models. The instance selection problem belongs to the NP-hard class, so it can be solved with an approximation tool. In this paper the firefly algorithm is proposed for solving the instance selection problem. This paper is one paper, where the firefly algorithm has been used to solve a discrete optimisation problem, when in more cases previously it has been used for solving continuous optimisation problems. The firefly-based instance selection algorithm is presented and its validation is carried out. The results of the computational experiment show that the algorithm is competitive with others. The results obtained are discussed and conclusions are formulated.

Automatic Identification System (AIS) is a telecommunication system created to allow ships to communicate with each other by exchanging messages containing information such as vessel’s ID, position, speed, heading, etc. AIS is useful in many cases, such as early ship collision detection. However, its terrestrial segment’s drawback is a relatively low range (about 74 km, roughly 40 nautical miles). Satellite Automatic Identification System (SAT-AIS) was introduced to overcome this limitation, but it suffers from its own problem known as packet collision. The satellite receives messages from multiple terrestrial cells and communication is synchronized within such cells, but not between them, thus messages got lost or damaged when they appear at the satellite at the same time. In this paper, we present a machine learning-based approach to reconstruct those missing messages and deeply investigate whether or not a density-based spatial clustering of applications with noise (DBSCAN) can be considered in the first stage of the reconstruction. The experiment focuses on findig the optimal parameters for the clustering, running it both on original and damaged data to finally ascertain that DSBCAN can distinguish individual trajectories in a dataset that can be further reconstructed.

The paper deals with the problem of supporting of a decision-making, prediction and decision validation in a human resource management system. In the paper a novel intelligent decision support system for human resource (HR) processes is discussed. The analytical path of the system has been designed and implemented, where machine learning algorithms have been proposed as basic tools for the monitoring different defined HR indicators. The paper includes the description of proposed approach and the discussion of selected experiments’ results.

The paper focuses on using the firefly algorithm for designing the radial-basis neural networks. The firefly algorithm belongs to a family of the global optimization tools. The firefly algorithm is a nature-inspired and one of the most powerful algorithms for solving the NP-hard optimization problems. In the paper the firefly algorithm is used as a tool for designing of the RBF network, including estimation of its output weights and transfer function parameters. The details of the implementation are discussed. Computational experiment has been carried-out with a view to investigate effectiveness of the discussed implementation. Experiment results have confirmed usefulness of the proposed approach. Conclusions include suggestions for future research.

The paper focuses on using stacking and rotation-based technique to improve performance and generalization ability of the machine learning classification with data reduction. The aim of data reduction technique is decreasing the quantity of information required to learn a high quality classifiers, especially when the data are huge. The paper shows that merging both stacking and rotation-based ensemble techniques with machine classification based on data reduction may bring additional benefits with respect to the accuracy of the classification process. The finding that has been confirmed by computational experiments. The paper includes the description of the approach and the discussion of the computational experiment results.

This volume comprises proceedings of the 2017 IEEE International Conference on INnovations in Intelligent SysTems and Applications (INISTA 2017), organized by Gdynia Maritime University, Poland, in cooperation with Yildiz Technical University, Turkey, IEEE Systems, Man, and Cybernetics Society (SMC), IEEE Poland Section, Poland Section of IEEE Systems, Man, and Cybernetics Society Chapter, IEEE SMC Technical Committee on Computational Collective Intelligence and Gdynia Maritime University's Students and Alumni Foundation. Technical Support of the Conference has been provided by the Poland Section of IEEE Computer Society Chapter. The INISTA 2017 took place in Gdynia, Poland, on July 3–5, 2017.

This paper focuses on the machine classification with data reduction. The aim of the data reduction techniques is decreasing the quantity of information required to learn a high quality classifiers. In this paper the data reduction is carried out by selection of relevant instances, called prototypes. To solve the machine classification problem with data reduction an agent-based population learning algorithm is proposed. The discussed approach bases on the assumption that the selection of prototypes is carried-out by a team of agents and that the prototype instances are selected from clusters of instances. The proposed procedure is called the stack generalization. It aims at improving the quality of the learning process and increasing the generalization capacity of the learning model. The paper includes the description of the approach and the discussion of the validating experiment results.

The article addresses the problem of feature selection in a kernel space and proposes the approach to selecting the most informative features for classification carried out using the RBF neural classifier. In the article, class-dependent and cluster-dependent methods for feature selection are considered. The process of feature selection is supported by the rotation-based ensemble technique. The feature selection problem is viewed as an optimization task solved by the agent-based population learning algorithm applied at the RBFN’s initialization and training stage. The proposed approach is validated experimentally, and the obtained results are compared with the results produced using other methods. Experiment results show that the proposed method of feature selection in a kernel space of the RBF neural networks can be considered as a useful approach to constructing high-quality RBFN-based classifiers.

The problem addressed in this study concerns mining data streams with concept drift. The goal of the article is to propose and validate a new approach to mining data streams with concept-drift using the ensemble classifier constructed from the one-class base classifiers. It is assumed that base classifiers of the proposed ensemble are induced from incoming chunks of the data stream. Each chunk consists of prototypes and information about whether the class prediction of these instances, carried-out at earlier steps, has been correct. Each data chunk can be updated by using the instance selection technique when new data arrive. When a new data chunk is formed, the ensemble model is also updated on the basis of weights assigned to each one-class classifier. In this article, two well-known instance-based learning algorithms—the CNN and the ENN—have been adopted to solve the one-class classification problems and, consequently, update the proposed classifier ensemble. The proposed approaches have been validated experimentally, and the computational experiment results are shown and discussed. The experiment results prove that the proposed approach using the ensemble classifier constructed from the one-class base classifiers with instance selection for chunk updating can outperform well-known approaches for data streams with concept drift.

The problem addressed in this paper concerns mining data streams with concept drift. The goal of the paper is to propose and validate a new approach to mining data streams with concept-drift using the ensemble classifier constructed from the one-class base classifiers. It is assumed that base classifiers of the proposed ensemble are induced from incoming chunks of the data stream. Each chunk consists of prototypes and can be updated using instance selection technique when a new data have arrived. When a new data chunk is formed, ensemble model is also updated on the basis of weights assigned to each one-class classifier. The proposed approach is validated experimentally.

This article proposes a novel approach to the radial basis function network (RBFN) design. Its main idea is to apply the agent-based population learning algorithm to the task of initialization and training RBFNs. The approach allows for an effective network initialization and estimation of its output weights. The initialization involves two stages, where in the first one initial clusters are produced using the similarity-based procedure and next, in the second stage, prototypes (centroids) from the thus-obtained clusters are selected. The agent-based population learning algorithm is used to select prototypes. In the proposed implementation of the algorithm, both tasks—RBFN initialization and RBFN training—are carried out by a team of agents executing various local search procedures and cooperating with a view to determine the solution to the RBFN design problem at hand. The performance of the RBFN constructed using the proposed agent-based approach is analyzed and evaluated. The proposed approach is also compared with different RBFN initialization and training procedures in the literature.

Clustering techniques have an influence on the quality and performance of RBF networks. The aim of the paper is to propose and evaluate a similarity-based approach for cluster initialization with the agent-based population learning algorithm used to select prototypes from the obtained clusters. The performance of the RBF network constructed using the proposed clustering and selection approach is analyzed and evaluated. The proposed approach is also compared with different RBF network initialization procedures.

The aim of this article is to propose and evaluate an agent-based population learning algorithm generating, through prototype selection, a representative training data set of the required size. The main feature of the proposed approach is selection of the representative feature vectors, called prototypes, from clusters of feature vectors constructed over the original training data set under the assumption that from each cluster a single prototype is obtained. Thus, the number of clusters produced from the original training data set has a direct influence on the size of the reduced data set. The process of selection is executed by a team of agents, which execute various local search procedures and cooperate to determine a solution to the instance reduction problem, aiming at obtaining a compact representation of the data set. Rules for agent cooperation during the clustering and selection processes are defined within the so-called working strategy used by the team of agents (A-Team) in question. The article proposes a set of procedures that are used by agents to produce clusters and select prototypes. The approach is validated experimentally using well-known benchmark data sets. In addition to the computational experiment used to validate the model, the article investigates the efficiency and performance of two different working strategies used by the proposed A-Team. Because the proposed approach is based on the population learning algorithm, which belongs to the class of the population-based methods, an evaluation of the influence of the population of solution size on the performance of the algorithm is also included.

Instance selection in the supervised machine learning, often referred to as the data reduction, aims at deciding which instances from the training set should be retained for further use during the learning process. Instance selection can result in increased capabilities and generalization properties of the learning model, shorter time of the learning process, or it can help in scaling up to large data sources. The paper proposes a cluster-based instance selection approach with the learning process executed by the team of agents and discusses its four variants. The basic assumption is that instance selection is carried out after the training data have been grouped into clusters. To validate the proposed approach and to investigate the influence of the clustering method used on the quality of the classification, the computational experiment has been carried out.

The paper deals with the distributed learning. Distributed learning from data is considered to be an important challenge faced by researchers and practice in the domain of the distributed data mining and distributed knowledge discovery from databases. An effective approach to learning from a geographically distributed data is to select, from the local databases, relevant local patterns, called also prototypes. Such a selection can be based on results of the data reduction process. The paper proposes to carry-out prototype selection at local sites in parallel, independently at each site, employing specialized software agents. To assure obtaining homogenous prototypes at a global level the consensus-based method is proposed and applied. The paper includes a detailed description of the proposed approach and a discussion of the computational experiment results.

The purpose of this article is to present the application of the A-team approach to solving some machine learning problems belonging to the supervised and unsupervised learning classes. Because the above problems are computationally hard, it is proposed to take advantage of the robustness and flexibility of population-based methods combined with the efficiency of multi-agent systems integrated within the A-team concept. The main part of the article summarizes the experiences of the authors gained while developing various A-teams and includes some examples of population-based multi-agent algorithms for solving problems from the machine learning domain. It can be concluded that population-based multi-agent algorithms can be competitive in comparison with other existing techniques for some machine learning problems.