Journal of Innovations in Computer Science and Engineering (JICSE)

Abstract This study presents an applied, multi-stage machine-learning pipeline for maternal health risk classification using the Maternal Health Risk dataset (1,014 samples, six clinical features, and three risk classes). Six classifiers were evaluated across four preprocessing stages: training-fold standardization alone, Edited Nearest Neighbors (ENN)-based cleaning, Synthetic Minority Oversampling Technique (SMOTE)-based class balancing, and sequential ENN-SMOTE preprocessing. Within each fold of 10-fold cross-validation, the scaler, ENN, SMOTE, model fitting, and ensemble construction were performed using only the training fold; the corresponding test fold remained untouched until evaluation. A fixed majority-voting ensemble comprising Random Forest, Extra Trees, and XGBoost was evaluated using class-wise, macro-averaged, and weighted-averaged accuracy, precision, and F1-score. In Stage 4, the hybrid ensemble achieved a weighted accuracy of 91.5% ± 3%, precision of 88% ± 3%, and F1-score of 87.2% ± 3%. The ensemble's precision for the high-risk class exceeded 90% across all four stages. These findings indicate competitive predictive performance on this dataset; however, direct comparison with previous studies is limited by differences in preprocessing methods, validation protocols, and model configurations. External and prospective validation is required before clinical use.

Abstract With the advent of new technologies, using various formats of digital gadgets is becoming widespread. In today's world, where everyday tasks are inevitable without technology, this extensive use of computers paves the way for malicious activity. Malware is one of the well-known and widely used means utilized for doing destructive activities by malicious attackers. Producing malware from scratch is somewhat difficult, so attackers tend to obfuscate existing malware and prepare it to become an unrecognizable program. Since creating new malware from an old one using obfuscation is a creative task, there are some drawbacks to identifying obfuscated malwares. In this research, we propose a solution to overcome this problem by converting the code to an image in the first step and then using a semi-supervised approach combined with contrastive learning. In this case, an obfuscation in the malware bytecode corresponds to an augmentation in the image. Hence, by utilizing meaningful augmentations, which simulate some obfuscation changes and combine them to generate complex ambiguity procedures, our proposed solution is able to construct, learn, and detect a wide range of obfuscations. This work addresses two issues: 1) malware classification despite the data deficiency and 2) obfuscated malware detection by training on non-obfuscated malwares. According to the results, the proposed method overcomes the data shortage problem in malware classification, as its accuracy is 90.1% when just 10% of data is used for training the model. Moreover, training on basic malwares without obfuscation achieved 96.21 percent accuracy in detecting obfuscated malware.

Abstract Population imaging studies rely on good-quality medical imagery before quantifying downstream images. This study provides an automated approach for image quality assessment (IQA) from cardiovascular magnetic resonance (CMR) imaging at scale. We identify four common CMR imaging artefacts: respiratory motion, cardiac motion, Gibbs ringing, and aliasing. Four datasets, including UK Biobank, York University (YU), the Universidad Carlos III (UCIII) and CMR-Tehran, were used to perform the experiments. This study proposes two deep-learning models for CMR IQA in spatial and frequency domains. The presented spatial-domain model also has domain adaptation. The accuracies of supervised 4-fold cross-validation experiments for UK Biobank, YU, UCIII and CMR-Tehran datasets are 99.41%, 75.78%, 89.46% and 67.87% for the spatial-domain and 87.46%, 63.76%, 80.25% and 58.48% for the frequency-domain. Domain adaptation results, considering UK Biobank as the source set and YU, UCIII and CMR-Tehran as the target sets, show the domain shift gap coverage between the datasets to the extent of +11.91%, +3.93% and +16.57%, respectively. Besides, by training and testing the spatial-domain model on 30,125 images from the UK Biobank, an accuracy of 89.56% was obtained in a training time of 394.80 seconds. Meanwhile, the frequency-domain model with training and testing on 180,750 images achieves an accuracy of 87.99% in a training time of 255.04 seconds. Thus, the frequency-domain model can achieve almost the same accuracy yet 1.548 times faster than the spatial model. The proposed models can detect four common CMR imaging artefacts by receiving images or the corresponding k-spaces.

Abstract Abstract— Sea Surface Temperature (SST) data reveal the temporal and spatial distribution of warm anticyclonic eddies and cold cyclonic eddies, impacting ocean behavior. The SST combined products attained adequate decisions to permit the recognition of mesoscale eddies with the introduction of altimeter operations and the availability of two or more altimeters at the same time. Climate change impacts ocean circulation and atmospheric anomalies linked to SST variations. Ocean eddies, vital for material and energy transport, require precise identification to advance oceanography. This study uses SST data from CMEMS in the Atlantic to introduce EddyNet, a deep-learning model for automatic eddy detection and classification. EddyNet's encoder-decoder architecture includes a pixel-wise classification layer, labeling each pixel as "0" (non-eddy), "1" (anticyclonic), or "2" (cyclonic). The high-resolution feature representation outperforms existing models, marking a significant leap in eddy detection accuracy and reliability. This study introduces EddyNet, a deep-learning model based on the U-Net architecture for automatic eddy detection and classification using Sea Surface Temperature (SST) data. The model was trained and evaluated on satellite imagery from the Copernicus Marine Environment Monitoring Service (CMEMS), achieving a training accuracy of 78.55%, a Dice score of 31.99%, and a precision of 0.9259. The recall values for different classes indicate that the model correctly identifies 99.51% of non-eddy pixels, 51.57% of anticyclonic eddy pixels, and 57.19% of cyclonic eddy pixels. These results demonstrate the effectiveness of deep learning in mesoscale eddy detection and highlight the potential for further optimization in classifying eddy structures with higher precision.

Abstract Abstract—Understanding individual motor signatures (IMS) is essential for personalized treatment and performance optimization. This study investigates the effectiveness of Fuzzy C-Means (FCM) clustering for identifying individual motor signatures from graphomotor tasks. We analyze various kinematic and geometric features, such as movement duration, velocity, and trajectory length, to reveal which aspects of motor behavior are most effective in distinguishing individuals. The results show that features like length of movement are particularly discriminative, while others, such as beta and velocity, offer weaker clustering outcomes.
Understanding individual motor signatures (IMS) is essential for personalized treatment and performance optimization. This study investigates the effectiveness of Fuzzy C-Means (FCM) clustering for identifying individual motor signatures from graphomotor tasks. We analyze various kinematic and geometric features, such as movement duration, velocity, and trajectory length, to reveal which aspects of motor behavior are most effective in distinguishing individuals. The results show that features like length of movement are particularly discriminative, while others, such as beta and velocity, offer weaker clustering outcomes.

Abstract Abstract—The study employed Deep Learning Frailty (DLF), a compelling neural modeling framework for predicting heart failure patient survival. The DLF embeds a notion of multiplicative frailty from classical survival analysis that deals with unobserved heterogeneity while exploiting the neural structure's strong capabilities in approximating any non-linear covariate relationship. The results showed that Incorporating frailty leads to significant improvements, and the DLF model performs better on average.

Abstract—The study employed Deep Learning Frailty (DLF), a compelling neural modeling framework for predicting heart failure patient survival. The DLF embeds a notion of multiplicative frailty from classical survival analysis that deals with unobserved heterogeneity while exploiting the neural structure's strong capabilities in approximating any non-linear covariate relationship. The results showed that Incorporating frailty leads to significant improvements, and the DLF model performs better on average.

Abstract—The study employed Deep Learning Frailty (DLF), a compelling neural modeling framework for predicting heart failure patient survival. The DLF embeds a notion of multiplicative frailty from classical survival analysis that deals with unobserved heterogeneity while exploiting the neural structure's strong capabilities in approximating any non-linear covariate relationship. The results showed that Incorporating frailty leads to significant improvements, and the DLF model performs better on average.

Abstract Abstract— Activation functions are essential for extracting meaningful relationships from real-world data in deep learning models. The design of activation functions is critical, as they directly influence the performance of these models. Nonlinear activation functions are commonly preferred since linear functions can limit a model’s learning capacity. Nonlinear activation functions can either have fixed parameters, which are predefined before training, or adjustable ones that modify during training. Fixed-parameter activation functions require the user to set the parameter values prior to model training. However, finding suitable parameters can be time-consuming and may slow down the convergence of the model. In this study, a novel fixed-parameter activation function is proposed and its performance is evaluated using benchmark MNIST datasets, demonstrating improvements in both accuracy and convergence speed.

Abstract— Activation functions are essential for extracting meaningful relationships from real-world data in deep learning models. The design of activation functions is critical, as they directly influence the performance of these models. Nonlinear activation functions are commonly preferred since linear functions can limit a model’s learning capacity. Nonlinear activation functions can either have fixed parameters, which are predefined before training, or adjustable ones that modify during training. Fixed-parameter activation functions require the user to set the parameter values prior to model training. However, finding suitable parameters can be time-consuming and may slow down the convergence of the model. In this study, a novel fixed-parameter activation function is proposed and its performance is evaluated using benchmark MNIST datasets, demonstrating improvements in both accuracy and convergence speed.

Abstract Abstract—Nowadays, advances in technology and medical science have led to significant changes in the field of healthcare. Consequently, an effort has been taken to develop an intelligent health assistant in the Persian language, focusing on the emergency department. To achieve this goal, a labeled dataset was prepared. Subsequently, an intelligent assistant architecture was developed, utilizing slot filling and speech act classification for natural language understanding. A dialogue manager was designed to address negation in patient statements, resulting in the classification of triage patients. Evaluation revealed that the assistant's performance matched that of emergency staff in 83% of cases.

Abstract—Nowadays, advances in technology and medical science have led to significant changes in the field of healthcare. Consequently, an effort has been taken to develop an intelligent health assistant in the Persian language, focusing on the emergency department. To achieve this goal, a labeled dataset was prepared. Subsequently, an intelligent assistant architecture was developed, utilizing slot filling and speech act classification for natural language understanding. A dialogue manager was designed to address negation in patient statements, resulting in the classification of triage patients. Evaluation revealed that the assistant's performance matched that of emergency staff in 83% of cases.

Abstract In recent years, one of the most widely studied areas in computer vision and natural language processing (NLP) is the interdisciplinary problem of Visual Question Answering (VQA), which involves the integration of computer vision and NLP.

Important challenges in this field include the need for large and suitable datasets as well as powerful hardware for training the model. Key factors to improve the performance of these models include selecting the appropriate neural network for processing the inputs, selecting the appropriate dataset, and the method of combining the features extracted from the inputs. Also, using different attention mechanisms can improve the overall performance of the system. Furthermore, incorporating various attention mechanisms into the model can significantly enhance the overall performance of VQA systems. In these systems, different neural networks are employed to process inputs: convolutional neural networks (CNNs) with various architectures are used for image processing, and different types of recurrent neural networks (RNNs) are used for text processing.

In this research, the architecture of the convolutional neural network is changed and the self-attention mechanism is used in text processing and the Skipgram language model is used for embedding the input text. The performance of the proposed model is evaluated on two datasets, VQA 1.0 and VQA 2.0. The results show that the proposed model has been able to increase the overall accuracy in the VQA 1.0 dataset to 67.25% and in the VQA 2.0 dataset to 61.57%, which show a significant improvement over the baseline models.

Abstract I. Abstract— This paper examines the legal frameworks governing the criminal liability of artificial intelligence (AI) in Iran and Saudi Arabia, focusing on how both countries address the evolving role of AI in criminal acts. With the rapid advancement of AI technologies—from narrow (weak) AI, which performs specific tasks, to general (strong) AI, capable of autonomous decision-making—complex legal and ethical questions have emerged. Specifically, this paper examines the applicability of three theoretical models of AI criminal liability: Perpetration-By-Another Liability, Natural-Probable-Consequence Liability, and Direct Liability.

II. The comparative analysis highlights that despite differences in legal traditions and societal contexts, both Iran and Saudi Arabia recognize that AI itself cannot bear criminal responsibility, and instead, liability is attributed to human actors, such as developers, users, and operators. The findings suggest that integrating technological progress with ethical and legal safeguards, grounded in Islamic jurisprudence, is essential for addressing the challenges posed by AI-related crimes in both jurisdictions.

Abstract Abstract—In today's world, DDoS attacks are becoming more common and complex; thus, they constitute a great challenge for network security under the auspices of SDN. The research effort described here proposes an integrated hybrid model called "LDA-ML," which leverages some state-of-the-art machine learning methods: LDA, naive bayes, random forest, and logistic regression. We optimize the data analysis process by leveraging LDA for feature selection and dimensionality reduction, followed by a sequential application of the classifiers to exploit their strengths. Evaluated on the CICDDoS-2019 dataset, the proposed model has achieved an outstanding accuracy of 98.98%, indicating the efficacy of the model in correctly classifying benign versus attack traffic. All of the above underlines the robustness of the proposed LDA-ML model, pointing to great potential for its application to continuously improve cybersecurity strategies against DDoS threats in SDN architectures. This holistic approach offers improvements in detection, while it also enriches diagnostic insights-an important contribution to finding effective security solutions in increasingly dynamic network environments.

Abstract Abstract— The issue of intrusion in security presents a fundamental challenge that can lead to serious damage in IT systems. Intrusion Detection Systems (IDS) serve as effective tools for identifying intrusion activities and generating alerts. However, traditional IDS methods often face issues such as low accuracy and long training times. Therefore, enhancing the performance and efficiency of these systems is crucial. The proposed approach in this study leverages evolutionary optimization algorithms combined with machine learning approaches to improve accuracy and training speed in IDS and better manage large volumes of data. This combination leads to the development of an Evolutionary Neural Network (ENN) that enhances and optimizes IDS performance. In this approach, BUZOA and Ant Colony Optimization (ACO) algorithms are used for feature selection, and decision tree, k-nearest neighbor, support vector machine, and deep neural network algorithms are used for classification and intrusion detection. The dataset used in this research is from the CICDDOS2019 database, containing 54,000 samples and 22 initial features. The experimental results indicate that among the metaheuristic algorithms BUZOA and ACO, and their combinations with decision tree, k-nearest neighbor, and support vector machine, the BUZOA-CNN hybrid algorithm with an average RMSE of 0.0117 and an accuracy of 96.32% performs better than other algorithms.

Abstract Abstract—The rapid advancement of artificial intelligence (AI) offers new avenues for enhancing the identification, control, and autonomous operation of marine vehicles. This study investigates the application of AI techniques in maritime environments, focusing on object detection, navigation, and autonomous control to support safe and efficient operations in various marine conditions. Key objectives include evaluating machine learning models for identifying and tracking marine vehicles and the development of intelligent control algorithms that can adapt to dynamic oceanic settings. Methods involved training convolutional neural networks (CNNs) on datasets of marine images for object identification and using reinforcement learning (RL) algorithms to optimize the control systems of autonomous marine vehicles. Results demonstrate that CNN-based models achieve high accuracy in vehicle identification, even under challenging visual conditions such as low lighting or occlusion. At the same time, RL-driven control systems adapt effectively to complex, fluctuating marine environments. Simulated and real-world testing indicated that these AI techniques improve vessel maneuverability and response times, leading to more efficient and safer operations. In conclusion, this study highlights the potential of AI to revolutionize marine vehicle identification and control, with implications for enhanced security, efficiency, and sustainability in maritime operations. It is advisable to conduct additional research to improve these models, enabling their application across a wider range of marine environments.

Abstract Abstract—This paper presents an intelligent non-player char- acter(NPC) for a survival game with a modified Q-learning-based scheme. Due to the dynamics of the computer game environment, reinforcement learning is employed to make this agent smart. This leads the agent to react appropriately based on the game’s scenario by choosing an action that provides a higher reward in the current situation. This is like a brain for the target NPC that processes different situations and reacts appropriately. Our intelligent agent is applied to a sample survival game with different complexity levels. In this game, multiple characters and objects alongside win-and-lose scenarios are considered. Our designed intelligent NPC is equipped with modified Q-learning to interact and try different actions on objects and learn about them. This learning process leads to an experience saved in the designed agent to react best to the environment. The efficiency of our proposed approach is evaluated through multiple scenarios and the appropriate reaction of the NPC is verified.

Abstract Abstract— This research proposes a novel approach to create a foundation for dynamic difficulty adjustment (DDA) within computer games that use procedural content generation (PCG), utilizing imitation learning to optimize gameplay. When PCG is used in creating the levels and enemies within a game, the difficulty adjustment must be ensured so that the game is not too hard or too easy for each player. However, PCG is random by nature, and thus, the developers may have a challenging task of adjusting the difficulty for each player in such games. The study aims to address these limitations by developing a foundation for DDA models based on imitation learning. The proposed model incorporates an imitation learning component, referred to as the 'Clone,' which replicates the player’s behavior, alongside an enemy creator agent that leverages procedural content generation (PCG) to design enemies. By analyzing the Clone's performance against these procedurally generated enemies, the system ensures the creation of fair and engaging levels. To this end, a 2D platformer Unity game using PCG was developed, and imitation learning was utilized through Unity's ML-agents module. These models were used to mimic the players' play-style to predict the player's performance in PCG-generated levels. Three separate models were created to mimic five players. It was observed that two of these models could mimic players' performance, showing that this method can be used to implement DDA.

Abstract Abstract— The brain is a complex organ that undergoes changes with age, and predicting brain age is crucial for monitoring brain health. It provides valuable insights into brain function and helps in the prevention of neurological diseases. This research predicts brain age through age classification based on fMRI data from the HCP dataset, consisting of individuals aged 22 to 36 years. After training a graph convolutional neural network, the model achieved an accuracy of 0.73 on the test data, demonstrating an improvement over previous studies on the same dataset. An evolutionary approach was then applied to optimize the selection of brain regions using a Genetic Algorithm to identify important and informative regions. This selection and optimization process maintained good predictive accuracy while reducing the number of brain regions. The results indicate that, despite using only half the original number of brain regions (8 regions), the model's accuracy remained at 0.65, showing only a slight decline. This highlights the significance of these regions in brain age classification. Identifying these key regions can contribute to the early diagnosis of brain and neurological diseases, enabling experts to better understand and manage the brain aging process.

Abstract Abstract— Deep learning has emerged as a transformative technology in ophthalmology, addressing critical challenges in the diagnosis and management of eye diseases such as diabetic retinopathy (DR), glaucoma, age-related macular degeneration (AMD), and central serous chorioretinopathy (CSCR). These conditions, among the leading causes of preventable blindness, require accurate and timely detection, which is often limited by traditional diagnostic methods due to inefficiency and the complexity of interpretation. The goal of this study is to examine the applications of deep learning in the diagnosis of ophthalmic diseases and to help researchers gain a better understanding of recent advances in model development, identify challenges associated with widespread implementation of these models in real-world applications, and outline future research directions in this area. Methodologically, recent studies using convolutional neural networks (CNNs), vision transformers, and hybrid models demonstrate high diagnostic accuracy and potential for early disease detection. Applications extend beyond disease diagnosis to lesion segmentation, disease progression monitoring, and personalized treatment planning. Deep learning systems have demonstrated comparable or superior diagnostic performance to human experts in detecting diseases such as DR and glaucoma. Despite these advances, challenges remain, including limited generalizability, data bias, and the need for explainable AI models to foster clinical trust and adoption. Addressing these challenges through improved model transparency, diverse datasets, and ethical frameworks will be critical to integrating deep learning into routine ophthalmic practice. This review highlights the significant advances in deep learning-driven ophthalmology and outlines a path for future research to optimize its clinical implementation.

Abstract Abstract— Voice Pathology Detection (VPD) aims to identify voice impairments through the analysis of speech signals, providing a foundation for developing diagnostic tools in advanced healthcare services to the public. This paper contributes to the development of efficient and accurate models based on deep learning (DL) for automatic VPD using sustained vowels of speech data. Therefore, this study explores the comparative efficacy of Mel-Frequency Cepstral Coefficients (MFCCs) and Linear Predictive Coding (LPC) as acoustic features extracted from vowels /i/, /a/, and /u/. Using the AVFAD database, we utilized and optimized a Convolutional Neural Network (CNN) as a DL model to classify healthy and pathological voices, prioritizing both accuracy and computational efficiency for real-time applications. Our findings reveal that 20 MFCC features extracted from vowel /i/ achieve the highest accuracy, with the optimal model reaching approximately 88% on test data. Our findings reveal that 20 MFCC features extracted from vowel /i/ achieve the highest accuracy, with the optimal model reaching approximately 88% on test data.

Abstract Abstract—Extracting organizational duties from legal documents is a critical yet challenging task, particularly in low-resource languages like Persian. This paper presents an innovative approach that integrates state-of-the-art Named Entity Recognition (NER) with advanced segmentation techniques and Large Language Models (LLMs) to accurately identify and extract duties assigned to organizations from Persian legal texts. Leveraging the power of the BERT-based model for NER, we enhance the recognition of relevant entities and ensure precise linkage to target organizations. Our method involves segmenting documents into sentences with an enhanced POS-based tokenizer, followed by the retrieval of contextually relevant segments based on the detected entities. We then explore the effectiveness of different LLM configurations, including a hierarchical approach that leverages both small and large models. Our experiments demonstrate that the hierarchical approach, combining ’Llama-3.1-8B’ and ’gpt-4o’, achieves an F1-score of 0.7901, significantly outperforming single-model approaches. This research underscores the potential of LLMs in legal text analysis, paving the way for more advanced tools in Natural Language Processing. Future work will include testing on a broader range of organizations, refining prompt engineering techniques, and enhancing model interpretability.

Abstract Abstract— Performance-based design of bridges requires prediction of different damage states of components. (RC) piers are key components in bridge system, which may experience severe damages during earthquakes. Therefore, seismic damage assessment of RC bridges depends strongly on predicting failure modes RC piers. Using machine learning for damage evaluation of structures is becoming increasingly popular in earthquake engineering. This study implements three different machine learning techniques to capture different damage limit states of RC bridge piers under seismic loading. For this purpose, three machine learning techniques including K-Nearest Neighbors (KNN), Artificial Neural Networks (ANNs) and decision tree regressions were utilized for predicting four damage states of a RC bridge piers tested experimentally under seismic excitations based on drift limits. The efficiency of the three algorithms in damage prediction of RC piers were compared.

The efficiency of the three algorithms in damage prediction of RC piers were compared. The efficiency of the three algorithms in damage prediction of RC piers were compared.

Abstract Abstract—Traditional Long Short-Term Memory (LSTM) networks are effective for handling sequential data but have limitations such as gradient vanishing and difficulty in capturing long-term dependencies, which can impact their performance in dynamic and risky environments like stock trading. To address these limitations, this study explores the usage of the newly introduced Extended Long Short-Term Memory (xLSTM) network in combination with a deep reinforcement learning (DRL) approach for automated stock trading. Our proposed method utilizes xLSTM networks in both actor and critic components, enabling effective handling of time series data and dynamic market environments. Proximal Policy Optimization (PPO), with its ability to balance exploration and exploitation, is employed to optimize the trading strategy. Experiments were conducted using financial data from major tech companies over a comprehensive timeline, demonstrating that the xLSTM-based model outperforms LSTM-based methods in key trading evaluation metrics, including cumulative return, average profitability per trade, maximum earning rate, maximum pullback, and Sharpe ratio. These findings mark the potential of xLSTM for enhancing DRL-based stock trading systems.

Abstract Abstract - - Abstract - - Abstract - This paper proposes a master-slave approach to simultaneously control two drones with the aim of carrying an object toward a goal. The proposed method utilizes the Double Deep Q-Learning (DDQN) technique to train a master agent to be able to carry an object toward a goal with the help of a slave agent. This procedure is implemented such that the master agent gathers the observations and specifies the actions to be made by itself and the slave agent. Indeed, the slave agent just applies a predefined action and does not process any input for producing the output. This manner of learning, leads to a unified convergence to an optimal solution compared to the situation in which each agent is trained separately. To verify the functionality of the proposed method, the algorithm is examined in the webots simulation environment. The simulations show that the introduced method has a good performance when controlling the drones to reach to the goal. The introduced method, other than algorithmic benefits which leads to a faster convergence of the model, suggests some reduction in the processing demand. The reason is that the learning procedure is guided by one of the agents and consequently only one of the agents is responsible for doing the calculations that lead to choosing the action. In this scenario, the slave agent does not require any processing resources for choosing the action and just simply applies a predefined action dictated by the master agent.

Abstract Abstract— With the increase in textual data generated on the internet and the limited time individuals have for reading, the need for automatic text summarization is more essential than ever. One application of summarization is title generation. The goal of this study, which falls within the field of digital humanities and interdisciplinary studies, is to provide a framework for title generation through extractive and abstractive summarization methods, focusing specifically on chapters of the Qur'an. For extractive summarization, eleven different methods have been examined, some of which are novel and innovative. For the abstractive part and title generation, several models have been trained to select the most effective one. In this research, the Persian translation of the Qur'an is used as the primary source, and a dataset was created based on the first ten parts (juz) of the Qur'an, including extractive summaries, abstractive summaries, and titles for various sections of the chapters. The results of this study indicate that the titles generated through summarization are close to human-generated titles, based on BERTScore, R-1, R-2, and R-l values of 21.03, 6.85, 20.73, and 52.51, respectively. It is important to note in the evaluation that a single fixed title does not exist for a document; multiple titles may also be valid. In human evaluation, we observed that the average score produced by the proposed approach is 0.59, while for the best results from other approaches, this value is 0.44.

Abstract Abstract—In today's world, simplifying compound and complex sentences into simple sentences is crucial for enhancing machine understanding in various natural language processing (NLP) tasks, such as inference, machine translation, and information extraction. This simplification process improves accuracy. Consequently, our research is inspired by a text simplification method called "split and rephrase." We introduce a new sequence-to-sequence text generation model that transforms complex sentences into simple ones based on the conjunction "and" in Persian. By utilizing linguistic models with millions or even billions of parameters, our approach facilitates a better understanding of text complexities and more accurate identification of breaking points. Our results show an output accuracy of 0.47 in the BLEU score for the generated simple sentences, which are both grammatically correct and fluent. By utilizing linguistic models with millions or even billions of parameters, our approach facilitates a better understanding of text complexities and more accurate identification of breaking points. Our results show an output accuracy of 0.47 in the BLEU score for the generated simple sentences, which are both grammatically correct and fluent.

Abstract Abstract— Telecommunications companies rely on recommendation systems to deliver personalized services and enhance customer satisfaction. Traditional methods, such as Collaborative Filtering (CF) and Content-Based Filtering (CBF), often fall short in capturing the complex relationships and social influences inherent in large telecom networks. In this paper, we propose a novel Graph Neural Network (GNN)-based recommendation system that integrates customer profiles with graph data representing customer interactions (e.g., calls, messages). The system uses the GraphSAGE architecture to aggregate information from each customer’s network, enabling it to learn from both direct and indirect relationships. By combining customer demographic and usage data with interaction networks, our model provides more accurate and personalized service recommendations.

We evaluate the system on a real-world telecom dataset, comparing it with traditional models, including CF, CBF, and Matrix Factorization (MF). The GNN-based system achieves a significant performance boost, with a precision of 0.81 and an F1-score of 0.80, outperforming all baselines. These results highlight the ability of GNNs to capture social and communication patterns, making them highly effective for telecom recommendations. Future work will explore the scalability of the system and its application to real-time data, further enhancing its potential for customer retention and revenue growth.