Software defect prediction (SDP) plays a critical role in improving software quality by identifying defective components during development and maintenance. This paper introduces FS-iSDP, an interpretable few-shot learning approach based on Siamese Neural Networks, designed to address the challenges posed by highly imbalanced SDP datasets. The proposed model learns to measure the similarity between software application classes represented through software metric-based features, enabling effective defect classification with limited training examples. Experiments were conducted on the Apache Calcite open-source project using a cross-version setup that simulates realistic software evolution. The results show that FS-iSDP achieves high recall and strong AUC values across multiple software versions, even as the number of defects decreases and class imbalance intensifies. To improve performance and reduce computational costs, Univariate Feature Selection is applied, which leads to improved precision and critical success index, along with a significant reduction in training time. In addition, a comparative evaluation is performed against state-of-the-art SDP methods, showing that FS-iSDP achieves competitive performance in most metrics. Finally, LIME analysis is used to interpret the model predictions, offering insight into which features most influence defect classification decisions. This approach proves effective for realistic, evolving software systems, and provides both predictive capabilities and interpretability in scenarios with scarce defect labels.

Medical Visual Question Answering (MVQA) is a crucial task at the intersection of multimodal machine learning and the medical domain. An MVQA system aims to obtain an answer from an input question and image and can have many applications, from facilitating healthcare to advancing education. Given the nature of the medical field, the development of MVQA models presents specific challenges, including data limitations and the need for explainable results. As a step towards addressing these problems, we propose MultI-MVQA, a multitask framework designed to enhance the performance of an MVQA system across various base models. Our proposal integrates additional annotated data, such as question, answer, and organ types, within a cascaded multitask architecture. By applying our MultI-MVQA framework to two base models from the MVQA literature, we achieve improved performance. Furthermore, to enhance model interpretability, we use the Local Interpretable Model-agnostic Explanations algorithm for both textual and visual inputs, and present an analysis of the explanations from both computational and biological perspectives.

In the last decade, the area of code generation based on natural language was one of the most studied machine learning topics. The paper addresses the problem of code generation from natural language, by generating a syntax-error-free generator model, which creates an Syntax-based model, later translated into code, for generating the structure of the code. Two approaches are comparatively investigated for generating the structure of the program. The first approach generates code templates in the form of an Abstract Syntax Tree, while the second generates the code in the form of an Abstract Syntax Graph, a new introduced concept which reduces the initial redundancy of Abstract Syntax Trees and uses it as a new way to generate code. The proposed methodology is tested on two literature data sets and on malware detection code generation based on a real data set containing MITRE ATT&CK techniques. The results outperform the state-of-the-art Abstract Syntax Tree approaches by 2.46% and the plain text-based approaches with more than 12.5%, highlighting that the proposed methodology learns better the structural representation than other literature approaches. • We propose two approaches for generating the structure of the code. • Abstract Syntax Trees and Abstract Syntax Graphs are proposed for code generation. • Experiments are performed on public data and on malware detection code generation. • Our results outperform the state-of-the-art Abstract Syntax Tree approaches by 2.46%. • The results also outperform the plain text-based related work with more than 12.5%.

The paper addresses the topic of radar echo temporal extrapolation which is of major interest in both operational and research meteorology. Weather radar measurements are an important data source used by operational meteorologists for weather analysis, radar refectivity having a significant influence on short-term heavy rainfall prediction. Thus, extrapolating radar products’ values is important for early storm evolution assessment. The paper proposes SepConv-ens approach for temporal extrapolation of radar observations using an ensemble of three separable convolution-based deep learning models. Experiments performed on real radar data from the Romanian National Meteorological Administration (NMA) highlight a good performance of SepConv-ens in predicting radar data up to more than 40 minutes ahead and a good correlation between the radar measurements and the predictions in terms of spatial and intensity evolution of the radar echoes. SepConv-ens is integrated in the operational visualization software utilized by the Romanian NMA and is the first attempt, at the national level, to offer an artificial intelligence-based automated assistance for operational meteorologists.

The Rock-Paper-Scissors game is a popular zero-sum game of cyclic nature, with a mixed-strategy Nash-equilibrium that has been the subject of a large number of studies and is of particular interest for economy, sociology and artificial intelligence.While there are numerous studies exploring evolutionary dynamics and learning, the overwhelming majority of these consider the game in its classical form, and two important axes with potential relevance remain unexplored.First, studies with policy-based reinforcement algorithms are lacking, and second, few existing investigations attempted to study such cyclic games with more than two players.The present work aims to address both of these matters.

In the context of new research in the software defect prediction (SDP) task using pre-trained language models, the present study aims to analyze the relevance of semantic features extracted using BERT-based language models for the detection of defective source codes. RoBERTa and CodeBERT-MLM language models are used to generate source code embeddings that capture semantic and contextual features which help in solving language understanding tasks such as SDP. The learned representations are then fed to a neural network-based SDP classifier in order to decide which of the learned code embeddings are more informative in discriminating between faulty and non-faulty software entities. Extensive experiments are conducted in a cross-version SDP scenario for Apache Calcite, an open-source framework for data management. The evaluation results of the defect classifiers show a statistically significant improvement when the code representations are learned by the pre-trained models compared with the semantic representations provided by other natural language-based models, doc2vec and LSI.

Software deffect prediction, a problem of major relevance within the search-based software engineering field, aims to enhance software quality by early and precisely uncovering faulty software modules. Accurate detection of software defects in new software releases might increase the performance of the software development process in terms of cost, time and software quality. Most approaches from the software deffect prediction literature try to develop general solutions that are designed to work with any type of software deffect. From a software engineering perspective, software defects may take various forms and identifying/fixing different types of defects requires different approaches. Starting from the assumption that specific types of software defects have a particular behaviour, we are introducing in this paper, as a proof of concept, an unsupervised learning-based methodology for mining behavioural patterns for specific classes of software defects and identifying features which would be relevant for detecting the uncovered classes. The experiments performed on an open-source software deffect prediction data set collected from all releases of the Apache Ivy software highlight that the results obtained by applying the proposed methodology are highly correlated with the way human domain experts categorise and address software defects. Creating software deffect prediction models that are specifically tailored for different software deffect types may improve the accuracy of the developed models, open the possibility to apply different sets of predictive models based on the domain of the software and may accelerate the adoption of software deffect prediction approaches by the industry.

Nowcasting represents a short-term weather forecast of how the atmospheric state will evolve during the next time period, typically less than two hours. It is vital for generating society-level emergency alerts in order to take timely actions and responses to potential disasters. The objective of the paper is to improve upon current nowcasting methods by applying a Deep Learning model that uses Convolutional Long-Short Term Memory Networks on a combination of satellite data. It is proposed a model ConvS Now for short-term prediction of satellite images that would be useful for precipitation nowcasting. The proposed model was trained and evaluated on satellite imagery collected by EUMESAT's Meteosat-11 satellite utilizing the Severe Storms RGB product. The experimental results performed a subset of the Meteosat-11 data spanning Europe demonstrate that this model can enhance weather short-term forecasting, reduce costs and time, and improve the general quality of predictions, as a normalized mean of absolute errors of 1.6% was attained, outperforming every other baseline approaches considered for comparison. A relative improvement of more than 30% has been achieved by the ConvS Now compared to the baselines, our proposed model being able to capture the spatio-temporal features of the weather evolution.

Software authorship identification has a major relevance in software development, as it can be of help in detecting potential problems with the source code, identifying false reporting of authorship within a project and it can contribute to the improvement of the code review and testing effort prioritization. This paper introduces SoftId, a one-class classification model. The model is trained to encode and recognize the programming style of a given set of software developers, thus gaining the ability to detect whether a certain source code is authored by a developer from the original set or it is an “unknown” software developer. The novelty of the SoftId classifier resides in solving the software authorship identification task in an open-set configuration using a deep autoencoder and based on textual representations of the source codes. The experimental results highlight a significant improvement of SoftId compared to One-class Support Vector Machine classifier used in the literature for anomaly and novelty detection.

Deep learning is in a continuous evolution and many domains benefit from this substantial progress in the development of intelligent solutions. While this progress has been swift, there are more and more applications and the specific requirements of each application domain entails much work done by researchers to adapt existing models or create new ones. This traditional approach has produced many successful designs, but recently, automated methods of finding neural network architectures emerged. We introduce IntelliSwAS approach for optimizing deep neural network architectures for a classification or regression task. Image classification was selected as the task, but IntelliSwAS is generic enough such that it could be successfully applied on other classification or regression problems. A particle swarm-based optimization algorithm is proposed for the automatic search for convolutional neural network architectures. The search technique is enhanced by a machine learning model (DAGRNN) which we designed for predicting the quality of the network architectures and thus increasing the performance of the algorithm. The proposed model is able to process data structured as directed acyclic graphs in general and we applied it specifically on network architectures. The network architecture that we discovered using IntelliSwAS surpassed 89.8% of the competing image classification models that we considered for comparison.

Authorship attribution is the task of determining the likely author of a given text, with applications in domains such as literature and literary history, social network analysis, software engineering and cybersecurity. AutoAt, a deep autoencoder-based classification model which exploits the ability of autoencoders to encode meaningful data patterns is proposed to solve this task. Experiments are conducted on a data set of 1571 poems authored by 8 Romanian poets using a distributed document representation. The proposed approach obtains comparable or better results with respect to other machine learning classifiers. Additionally, the formulation of the AutoAt model allows for the computation of the probability that a test instance belongs to a given author class, which may be a useful property in a variety of authorship attribution applications. This aspect and the fact that AutoAt performs well in the difficult task of authorship attribution on poetic data without the step of feature engineering being informed by domain knowledge show that the proposed classifier is a general one, with potential to be used successfully in other fields.

Due to the increasing number of severe phenomena in many regions of the world, weather nowcasting, which is the weather forecast for a short time period, is one of the most challenging topics in meteorology. The weather radar and satellite are essential tools currently used by operational meteorologists for nowcasting. Issuing nowcasting warnings based on radar and satellite data is a complex task, due to the large volume of data that should be analyzed by meteorologists in order to make decisions. We are introducing in this paper DeePS at, a convolutional neural network architecture for short-term satellite images prediction that would be useful for weather nowcasting. The experimental evaluation is conducted on satellite data collected by EUMETSAT’s Meteosat-11 satellite, using five satellite products. The obtained results are analyzed and compared to the results of similar approaches. An average normalized mean of absolute errors of 3.84% was obtained for all satellite products, highlighting this way the effectiveness of DeePS at model.

Imbalanced classification represents a challenge for supervised learning, as an unequal distribution of classes in the training data set is mainly connected to poor predictive performance for the minority class. However, usually the minority class is the most relevant one, from a practical perspective. But due to the imbalancement of the training data, the classification errors for the minority class are higher, as the classifiers are usually biased to predict the majority class. In this paper we investigate the use of autoencoders for improving the predictive performance for imbalanced binary classification problems. As an application domain we consider breast cancer detection, that is an imbalanced classification problem of great interest in the medical domain. According to the World Health Organisation, breast cancer represents the primary cause of cancer mortality in women. Nowadays there is an increasing interest in applying conventional machine learning and more recently deep learning techniques in the breast cancer detection field by helping medical experts in the early detection of the disease. One of the paper’s goal is to investigate the ability of deep autoencoders to learn patterns within the classes of benign and malignant instances. Secondly, we propose and compare two autoencoders-based classification models for breast cancer detection. The performances of the proposed models were empirically assessed on data sets previously used in the breast cancer detection literature. The results show that our best model compares favourably with the results of most of the classifiers used for comparison and that it is able to handle well the data imbalancement.

The paper approaches the topic of nowcasting, one of the hottest topics in meteorology which deals with the problem of short-term forecasting of severe weather phenomena. Various types of meteorological data, including radar measurements, satellite data and weather stations’ observations are currently used for forecasting severe weather events. Radar data is one of the important sources used by meteorologists for nowcasting and for providing alerts for severe weather events. We are proposing a new one-class classifier, named RadRAR (Radar products’ values prediction using Relational Association Rules) for convective storms nowcasting based on radar data. More specifically, RadRAR is trained on radar data collected from normal weather conditions and learns to predict whether the radar echo values will be higher than 35dBZ, i.e. likely to indicate the occurrence of a storm. RadRAR is intended to be a proof a concept that relational association rule mining applied on radar data is helpful in discriminating between severe and normal weather conditions. Real radar data provided by the Romanian National Meteorological Administration is used for evaluating the effectiveness of RadRAR. A Critical Success Index of 61% was obtained, outperforming similar approaches from the literature and highlighting a good performance of RadRAR.

Identifying defective software components is an essential activity during software development which contributes to continuously improving the software quality. Since relatively numerous defects are due to violated software dependencies, coupling metrics could increase the performance of software defect prediction. Among various measures expressing the coupling between software components, the conceptual coupling metrics capture similarities based on the semantic information contained in the source code. We are introducing a new conceptual coupling based metric suite, named COMET, for software defect prediction. Experiments conducted on publicly available data sets, using both unsupervised and supervised learning models, emphasize that COMET metrics suite is superior to the software metrics widely used in the defect prediction literature.

This paper analyses the problem of predicting students’ academic performance, a subject that is increasingly investigated within the Educational Data Mining literature. For a better understanding of the educational related phenomena, there is a continuous interest in applying supervised and unsupervised learning methods for obtaining additional insights into the students’ learning process. The problem of predicting if a student will pass or fail at a certain academic discipline based on the students’ grades received during the semester is a difficult one, highly dependent on various conditions such as the course, the number of examinations during the semester, the instructors and their exigences. We propose a new classification model, S PRAR (Students Performance prediction using Relational Association Rules) for predicting the final result of a student at a certain academic discipline using relational association rules (RARs). RARs extend the classical association rules for expressing various relationships between data attributes. Experiments are performed on three real academic data sets collected from Babeş-Bolyai University from Romania. The performance of the S PRAR classifier on the considered case studies is compared against existing related work, being superior to previously proposed students’ performance predictors.

Predicting weather, and particularly severe weather, is an important challenge both for meteorological and machine learning researchers. The complexity and difficulty of the problem is mainly due to the chaotic character of the atmosphere and the implicit large set of meteorological information (radar, satellite or ground meteorological observations) which have to be analyzed by meteorologists. Thus, understanding the relationships between various meteorological parameters extracted from radar observations may be useful for providing additional comprehension about severe weather development and would help to identify situations when severe weather can occur. Self-organizing maps are being explored as an unsupervised classification model for detecting patterns in radar data which are relevant in predicting short-term weather changes. Experiments are performed on real radar data provided by the Romanian National Meteorological Administration. With the main goal of analyzing how the values for the weather radar products are evolving between consecutive radar scans, we empirically show that in general there is a slow change in the values over time, except for the situations when certain severe phenomena occur. The study conducted in this paper is aimed to provide a better insight regarding how the values of weather radar products are evolving in time both in calm and severe weather conditions, with the broader goal of using these findings for weather nowcasting.

Deploying fleets of mobile robots in real scenarios and environments raises several scientific challenges. One of them concerns the ability of the robots to adapt to the dynamics of their environment. We introduce DynFloR , a dynamic network flow based approach for finding optimal policies for data delivery in multi-robot network patrolling where the robots can communicate instantly and free of charge one to another when they meet, there is a periodicity of the robot meetings and the distribution of the data collected during the patrol is regular. Experiments on randomly generated synthetic examples are performed for evaluating the performance of the DynFloR method. The performed experiments empirically show that independent of the problem setting (such as number of robots, memory of the robots) the amount of data transferred to a base station per unit of time converges to an equilibrium state. The case of lost data has been also examined through various experiments, but it requires further experimentation as well as in-depth analysis.

Relational Association Rules (RARs) capture generic relations between attributes values in possibly large data sets. Due to their ability to uncover underlying semantically relevant patterns, they are of particular interest in data mining research and applicable in both unsupervised and supervised learning scenarios. With the aim of increasing the stability and expressiveness of the classical, non-gradual RARs, Gradual Relational Association Rules (GRARs) have been introduced. By generalizing the boolean relations to gradual relations, GRARs also capture the degrees to which generic relations are satisfied. In the current paper we introduce a new approach called DynGRAR (Dynamic Gradual Relational Association Rules Miner) for uncovering interesting GRARs in dynamic data sets which are incrementally extended with both new data instances and new data attributes. DynGRAR dynamically adjusts the set of all interesting GRARs. Through multiple experiments performed on publicly available software defect prediction data sets, we have evaluated DynGRAR versus applying the standard GRARs mining algorithm from scratch on the extended data. The results obtained emphasize the superior performance of the dynamic approach we propose.

Online data mining techniques are used to uncover relevant patterns in complex data which are dynamic by nature and thus continuously extended with real-time arriving data streams. Relational association rules (RARs), a data analysis and mining concept, extend the classical association rules so as to capture different relations between the attributes characterizing the data. This paper introduces a new Incremental Relational Association Rule Mining (IRARM) approach with the aim of progressively adapting the interesting relational association rules identified in a data set, when it is enlarged with new instances. We have experimentally evaluated IRARM on publicly available data sets. The reduction in mining time when using IRARM against mining from scratch emphasizes its efficiency in adapting the rules to real-time data extension.

Proteins are complex macromolecules which contribute to maintaining cellular environments and thus have fundamental roles in biological processes of living organisms. Understanding the conformational transitions of proteins represents an important stage towards comprehending protein function and would help to identify situations when mutations can occur. In this paper we use clustering as an unsupervised classification method in order to study the relevance of the residues’ relative solvent accessibility (RSA) values to analyze protein internal transitions. With the main goal of studying the evolution of RSA values between conformational transitions, we experimentally show that RSA values are slowly modifying as the protein undergoes conformational changes. The study conducted in this paper is aimed to provide a better apprehension of how proteins’ conformational transitions are evolving in time, with the broader goal of better understanding protein internal dynamics.