Vol. 42 · No. 3

Natural language processing (NLP) involves the computer analysis and processing of human languages using a variety of techniques aimed at adapting various tasks or computer programs to linguistically process natural language. Currently, NLP is increasingly applied to a wide range of real-world problems. These tasks can vary from extracting meaningful information from unstructured data, analyzing sentiment, translating text between languages, to generating human-level text autonomously. The goal of this study is to employ transformer-based natural language models to generate high-quality business names. Specifically, this work investigates whether larger models, which require more training time, yield better results for generating relatively short texts, such as business names. To achieve this, we utilize different transformer architectures, including both freely available and proprietary models, and compare their performance. Our dataset comprises 250 928 observations of business names. Based on the perplexity metric, the top-performing model in our study is the GPT2-Medium model. However, our findings reveal a discrepancy between human evaluation and perplexity-based assessment. According to human evaluation, the best results are obtained using the GPT-Neo-1.3B model. Interestingly, the larger GPT-Neo-2.7B model yields poorer results, with its performance not being statistically different from that of the GPT-Neo-125M model, which is 20 times smaller.

Pursuit-Evasion Game (PEG) can be defined as a set of agents known as pursuers, which cooperate with the aim forming dynamic coalitions to capture dynamic evader agents, while the evaders try to avoid this capture by moving in the environment according to specific velocities. The factor of capturing time was treated by various studies before, but remain the powerful tools used to satisfy this factor object of research. To improve the capturing time factor we proposed in this work a novel online decentralized coalition formation algorithm equipped with Convolutional Neural Network (CNN) and based on the Iterated Elimination of Dominated Strategies (IEDS). The coalition is formed such that the pursuer should learn at each iteration the approximator formation achieving the capture in the shortest time. The pursuer's learning process depends on the features extracted by CNN at each iteration. The proposed supervised technique is compared through simulation, with the IEDS algorithm, AGR algorithm. Simulation results show that the proposed learning technique outperform the IEDS algorithm and the AGR algorithm with respect to the learning time which represents an important factor in a chasing game.

Crime is hard to anticipate since it occurs at random and can occur anywhere at any moment, making it a difficult issue for any society to address. By analyzing and comparing eight known prediction models: Naive Bayes, Stacking, Random Forest, Lazy:IBK, Bagging, Support Vector Machine, Convolutional Neural Network, and Locally Weighted Learning – this study proposed an improved deep learning crime prediction model using convolutional neural networks and the xgboost algorithm to predict crime. The major goal of this research is to provide an improved crime prediction model based on previous criminal records. Using the Boston crime dataset, where our larceny crime dataset was extracted, exploratory data analysis (EDA) is used to uncover patterns and explain trends in crimes. The performance of the proposed model on the basis of accuracy, recall, and f-measure was 100

This paper introduces a model for the authentication of large-scale images. The crucial element of the proposed model is the optimized Pulse Coupled Neural Network. This neural network generates position matrices based on which the embedding of authentication data into cover images is applied. Emphasis is placed on the minimalization of the stego image entropy change. Stego image entropy is consequently compared with the reference entropy of the cover image. The security of the suggested solution is granted by the neural network weights initialized with a steganographic key and by the encryption of accompanying steganographic data using the AES-256 algorithm. The integrity of the images is verified through the SHA-256 hash function. The integration of the accompanying and authentication data directly into the stego image and the authentication of the large images are the main contributions of the work.

Attribute-based access control (ABAC) has higher flexibility and better scalability than traditional access control and can be used for fine-grained access control of large-scale information systems. Although ABAC can depict a dynamic, complex access control policy, it is costly, tedious, and error-prone to manually define. Therefore, it is worth studying how to construct an ABAC policy efficiently and accurately. This paper proposes an ABAC policy generation approach based on the CatBoost algorithm to automatically learn policies from historical access logs. First, we perform a weighted reconstruction of the attributes for the policy to be mined. Second, we provide an ABAC rule extraction algorithm, rule pruning algorithm, and rule optimization algorithm, among which the rule pruning and rule optimization algorithms are used to improve the accuracy of the generated policies. In addition, we present a new policy quality indicator to measure the accuracy and simplicity of the generated policies. Finally, the results of an experiment conducted to validate the approach verify its feasibility and effectiveness.

Sentiment classification plays a pivotal role in natural language processing (NLP), and prior research has established the efficacy of utilizing convolutional neural networks (CNNs) and long short-term memory (LSTM) in this task. However, these approaches suffer from individual performance limitations: CNNs are limited to extracting local information and fail to express context information adequately, while LSTM networks excel at extracting context dependencies but exhibit long training times. To address this issue, we propose a novel text classification algorithm based on a hybrid CNN-LSTM model that leverages the strengths of both approaches and overcomes their limitations by combining them. Our approach is evaluated on the IMDB dataset, and we present a hyperparameter optimization framework utilizing Random Search to increase the likelihood of producing an optimally performing model.

The domains of a protein provide an insight on the functions that the protein can perform. Delineation of proteins using high-throughput experimental methods is difficult and a time-consuming task. Template-free and sequence-based computational methods that mainly rely on machine learning techniques can be used. However, some of the drawbacks of computational methods are low accuracy and their limitation in predicting different types of multi-domain proteins. Biological language modeling and deep learning techniques can be useful in such situations. In this study, we propose BERTDom for segmenting protein sequences. BERTDOM uses BERT for feature representation and stacked bi-directional long short term memory for classification. We pre-train BERT from scratch on a corpus of protein sequences obtained from UniProt knowledge base with reference clusters. For comparison, we also used two other deep learning architectures: LSTM and feed-forward neural networks. We also experimented with protein-to-vector (Pro2Vec) feature representation that uses word2vec to encode protein bio-words. For testing, three other bench-marked datasets were used. The experimental results on benchmarks datasets show that BERTDom produces the best F-score as compared to other template-based and template-free protein domain boundary prediction methods. Employing deep learning architectures can significantly improve domain boundary prediction. Furthermore, BERT used extensively in NLP for feature representation, has shown promising results when used for encoding bio-words. The code is available at https://github.com/maryam988/BERTDom-Code.

Multidimensional hierarchical (mTree) data are very common in daily life and scientific research. However, mTree data exploration is a laborious and time-consuming process due to its structural complexity and large dimension combination space. To address this problem, we present mTreeIllustrator, a mixed-initiative framework for exploratory analysis of multidimensional hierarchical data with faceted visualizations. First, we propose a recommendation pipeline for the automatic selection and visual representation of important subspaces of mTree data. Furthermore, we design a visual framework and an interaction schema to couple automatic recommendations with human specifications to facilitate progressive exploratory analysis. Comparative experiments and user studies demonstrate the usability and effectiveness of our framework.

We propose a software layered architecture for autonomous vehicles whose efficiency is driven by pull-based acquisition of sensor data. This multi-process software architecture, to be embedded into the control loop of these vehicles, includes a Belief-Desire-Intention agent that can consistently assist the achievement of intentions. Since driving on roads implies huge dynamic considerations, we tackle both reactivity and context awareness considerations on the execution loop of the vehicle. While the proposed architecture gradually offers 4 levels of reactivity, from arch-reflex to the deep modification of the previously built execution plan, the observation module concurrently exploits noise filtering and introduces frequency control to allow symbolic feature extraction while both fuzzy and first order logic management are used to enforce consistency and certainty over the context information properties. The presented use-case, the daily delivery of a network of pharmacy offices by an autonomous vehicle taking into account contextual (spatio-temporal) traffic features, shows the efficiency and the modularity of the architecture, as well as the scalability of the reaction levels.

Brain-computer interfaces (BCIs) have demonstrated immense potential in aiding stroke patients during their physical rehabilitation journey. By reshaping the neural circuits connecting the patient's brain and limbs, these interfaces contribute to the restoration of motor functions, ultimately leading to a significant improvement in the patient's overall quality of life. However, the current BCI primarily relies on Electroencephalogram (EEG) motor imagery (MI), which has relatively coarse recognition granularity and struggles to accurately recognize specific hand movements. To address this limitation, this paper proposes a hybrid BCI framework based on Electroencephalogram and Electromyography (EEG-EMG). The framework utilizes a combination of techniques: decoding EEG by using Graph Convolutional LSTM Networks (GCN-LSTM) to recognize the subject's motion intention, and decoding EMG by using a convolutional neural network (CNN) to accurately identify hand movements. In EEG decoding, the correlation between channels is calculated using Standardized Permutation Mutual Information (SPMI), and the decoding process is further explained by analyzing the correlation matrix. In EMG decoding, experiments are conducted on two task paradigms, both achieving promising results. The proposed framework is validated using the publicly available WAL-EEG-GAL (Wearable interfaces for hand function recovery Electroencephalography Grasp-And-Lift) dataset, where the average classification accuracies of EEG and EMG are 0.892 and 0.954, respectively. This research aims to establish an efficient and user-friendly EEG-EMG hybrid BCI, thereby facilitating the hand rehabilitation training of stroke patients.

In order to perfectly meet the needs of business leaders, decision-makers have resorted to the integration of external sources (such as Linked Open Data) in the decision-making system in order to enrich their existing data warehouses with new concepts contributing to bring added value to their organizations, enhance its productivity and retain its customers. However, the traditional data warehouse environment is not suitable to support external Big Data. To deal with this new challenge, several researches are oriented towards the direct conversion of classical relational data warehouse to a columnar NoSQL data warehouse, whereas the existing advanced works based on clustering algorithms are very limited and have several shortcomings. In this context, our paper proposes a new solution that conceives an optimized columnar data warehouse based on CLARANS clustering algorithm that has proven its effectiveness in generating optimal column families. Experimental results improve the validity of our system by performing a detailed comparative study between the existing advanced approaches and our proposed optimized method.