Journal of Innovations in Computer Science and Engineering (JICSE)

Abstract Abstract - As artificial intelligence (AI) increasingly becomes central to digital transformation, businesses across industries are recognizing its transformative potential for enhancing efficiency, accuracy, and innovation. This article examines a structured framework for AI integration that empowers businesses to manage the challenges associated with AI implementation, emphasizing both technical and "soft" competencies crucial for successful implementation. Through a phased approach, including Discovery, Roadmap Design, Implementation, and Evaluation, this review provides actionable insights to align AI solutions with business objectives, optimize resources, and overcome organizational barriers. The framework highlights how AI-driven tools, such as predictive analytics, data mining, and automated decision-making systems, enhance strategic capabilities, streamline operations, and improve customer engagement. To ensure long-term success, this study underscores the significance of cultivating an environment that promotes innovation and teamwork. AI adoption requires not only robust data infrastructure and technical expertise but also strategic foresight, cross-functional collaboration, and a commitment to iterative learning. By integrating technical and soft knowledge, organizations can overcome challenges in AI adoption, such as resistance to change and uncertain ROI, by fostering a supportive environment that enables AI-driven growth. This article provides decision-makers with a thorough guide, equipping them with the insight needed to maximize AI’s potential for long-term competitive success in an evolving digital world.

Abstract Abstract—The paper aimed to analyze the thematic evolution and various networks of intellectual structures in the field of biomedical ontologies during 2014-2023. This applied research used an analytical and descriptive method, co-word techniques, and social network analysis. A web-based interface of bibliometrix, Microsoft Excel, and VOSviewer software were used for descriptive bibliometric study, data analysis, and network structure visualization. The period from mid-2020 to mid-2021 presented an increased dissemination of significant and prominent keywords within the overlay network in the field. Five major topic clusters were identified based on a co-occurrence network. These clusters labeled ‘gene ontology’, ‘biomedical informatics focusing on AI techniques’, ‘bioinformatics applications in biomarker discovery’, ‘protein interaction networks in Alzheimer's proteomics’, and ‘network-based molecular mechanism’. Basic clusters were ’gene ontology’, ‘bioinformatics’, and ‘gene expression’. Moreover, five clusters experienced significant developments between 2023 and 2024, namely ‘bioinformatics’, ‘deep learning’, ‘machine learning’, ‘transcriptome’, and ‘network pharmacology’. These topics are the latest and hottest concepts in this field. Clusters, namely ‘deep learning’,’ machine learning, and ‘ontology’ were recognized as niche and the most well-developed themes. The most mature and mainstream thematic clusters were namely ‘transcriptome’, ’prognosis’, and ‘rna-seq’. The most undeveloped and chaotic themes were ‘network pharmacology’ and ‘molecular docking’.

Abstract Parkinson's Disease (PD) is a prevalent neurological disorder marked by motor symptoms such as rigidity and tremors. Accurate and timely assessment of disease severity is essential for judging the efficacy of various treatment interventions. This study presents an innovative approach that employs computer vision technology paired with advanced deep learning techniques to enable precise evaluations of Parkinson's severity.

Leveraging the high accuracy of Kinect cameras in capturing essential movement patterns, our proposed system employs advanced convolutional neural networks, specifically incorporating mechanisms from the Spatial-Temporal Graph Convolutional Network (STGCN) and the Two-Stream Adaptive Graph Convolutional Network (2SAGCN). These architectures are adept at detecting movement anomalies and generating precise quantitative severity measures. To further enhance the performance of the 2SAGCN, we introduce distinct temporal and spatial attention modules, resulting in improved classification outcomes. The model achieves outstanding metrics, with accuracy, precision,recall,and F1 score recorded at 94.14 ± 0.26, 98.1 ± 0.12, 98.6 ± 0.05,and98.2 ± 0.02, respectively The severity classification framework distinguishes between11specific classes of Parkinson's symptoms, which are derived from 9 distinct motion categories.Within this framework, class 0 represents healthy individuals, while classes 0 to 1 correspond to varying degrees of severity in Parkinson's symptoms, resulting in a comprehensive classification system encompassing 99 distinct outcomes.

To further enhance the model’s accuracy, we have implemented strategies such as transfer learning and data 3D augmentation. This research marks a significant step forward in the realm of non-invasive, quantitative assessments of Parkinson's Disease, showcasing the potential of cutting-edge technology and state-of-the-art neural network architectures.

Abstract Abstract— This study examines the impact of using artificial intelligence chatbots on improving the quality of life for women with multiple sclerosis (MS) in Iran. Using a qualitative method and semi-structured interviews, the experiences of women participating in relation to the functionality of AI chatbots were analyzed. The findings indicate that chatbots can play a significant role as supportive and informational tools in managing the disease, reducing anxiety, and improving communication for these women. Additionally, these tools assist in organizing daily tasks and reducing feelings of loneliness. Although some participants pointed to an excessive reliance on chatbots, overall, the results show more positive effects compared to the disadvantages of this technology. Ultimately, the research suggests that future studies should explore the psychological and ethical impacts of using chatbots more deeply. Ultimately, the research suggests that future studies should explore the psychological and ethical impacts of using chatbots more deeply. Ultimately, the research suggests that future studies should explore the psychological and ethical impacts of using chatbots more deeply.

Abstract Structural reliability analysis (SRA) is associated with complex calculations and large number of simulations. In this paper, machine learning (ML) methods are integrated with SRA to reduce the overall intricacy and computational cost of direct SRA methods, such as the Monte Carlo simulation (MCS) method. An SRA is conducted in this paper on post-tensioned concrete members under the influence of prestress loss, and their reliability indices are obtained through the MCS method. The results of the SRA are used to create a database for data fitting of the ML algorithms. The algorithms are compared to find the most accurate ML model to be applied on the problem at hand. For the SRA, different stochastic parameters with specified probabilistic distributions are considered for the numerical models, and nonlinear dynamic analyses are conducted on them. Using the labeled data resulted from the SRA, five ML algorithms are compared; (i) linear regression, (ii) random forest, (iii) artificial neural network, (iv) k-nearest neighbors, (v) extreme gradient boosting. R-squared and root mean squared error are considered as the metrics used for the comparison of the ML models. Bayesian search is used for hyperparameter optimization of algorithms. The performance of the linear regression algorithm (R2=0.67 and RMSE=0.26) indicates that the SRA problems are highly nonlinear and linear algorithms cannot precisely map the relationships in data. However, the results show that extreme gradient boosting has the finest accuracy with R2=0.9 and RMSE=0.04. Additionally, its predicted values mostly have relative errors of less than ±30%.

Abstract Abstract— Intent detection and slot filling are crucial for understanding human language and are essential for creating intelligent virtual assistants, chatbots, and other interactive systems that interpret user queries accurately. Recent advancements, especially in transformer-based architectures and large language models (LLMs), have significantly improved the effectiveness of intent detection and slot filling. This paper, proposes a method for effectively utilizing low volume fine-tuning data samples to enhance the natural language comprehension of lightweight language models, yielding a nimble and efficient approach. Our approach involves augmenting new data while increasing model layers to enhance understanding of desired intents and slots. We explored various synonym replacement methods and prompt-generated data samples created by large language models. To prevent semantic meaning disturbance, we established a lexical retention list containing non-O slots to preserve the sentence's core meaning. This strategy enhances the model's slot precision, recall, F1-score, and exact match metrics by 1.41%, 1.8%, 1.61%, and 3.81%, respectively, compared to not using it. The impact of increasing model layers was studied under different layer arrangement scenarios. Our results show that our proposed solution outperforms the baseline by 10.95% and 4.89% in exact match and slot F1-score evaluation metrics.

Abstract Emotion recognition in text is a growing area in Natural Language Processing ( NLP ), essential for improving human-computer interactions by allowing systems to interpret emotional expressions. While much progress has been made in English, Persian emotion recognition has seen limited devel- opment due to resource constraints and linguistic challenges. In this study, we address these gaps by leveraging two key Persian datasets, ArmanEmo and ShortEmo, to train an efficient emotion recognition model. Using FaBERT, a BERT-based model optimized for Persian, we employ interme- diate fine-tuning on a large collection of informal and formal Persian texts to enhance the model’s adaptability to colloquial language. This step significantly improves comprehension of Persian text
variations, as reflected in reduced perplexity scores. Our final evaluations, incorporating accuracy, precision, recall, and F1 score metrics, demonstrate that this fine-tuned FaBERT model achieves strong performance in emotion recognition, providing a promising approach for NLP in low-resource languages...

Abstract Skin cancer is one of the deadliest but most prevalent types of cancer; as such, early diagnosis is urgently required to improve patient outcomes. This work presents a collaborative deep learning model that classifies skin cancer with respect to three different networks: EfficientnetB1, EfficientnetB2, and EfficientnetV2s on dermoscopic images. The proposed collaborative model has a multi-head attention mechanism, ensuring that this model has a better attention capability for improving its accuracy in the task of classification. The HAM10k dataset provided the proposed model with a platform for fine tuning with transfer learning, along with some augmentation techniques to handle class imbalance challenges and feature variations of lesions. The results for the ensemble model combined with Multi-Head Attention were very high: an accuracy of 97.11%, and precision, recall, and F1-score are also very high. These findings prove that our approach can dramatically improve automation in skin cancer detection. Therefore, it will be helpful in clinical dermatology for early diagnosis in medicine.

Abstract Abstract— Detecting white blood cells (WBC) in microscopic images is essential in medical diagnosis. Manual analysis of these images is time-consuming and has a high error rate. Using object detection for WBCs detection with deep convolutional neural networks (CNN) can be considered a practical and effective solution. In this study, a CNN model is proposed to classify these images. In order to achieve optimal training performance, CNNs have many hyperparameters, such as dropout rate, number of hidden units in each hidden layer, activation function, loss function and optimizer, which need to be optimized. Therefore, a hyperparameter optimization approach based on a genetic algorithm is suggested, which can then be used to select the best combination parameters to improve accuracy and efficiency in detecting white blood cells in microscopic images. This new approach is significant and flexible for medical technicians to use in clinical practice for examining blood cell microscopy. In this research, the images were classified into five classes and the mean accuracy of the model for the five classes was 87%, which is considered a good accuracy for classification into five classes.

Abstract Understanding dynamic systems like disease outbreaks, social influence, and information diffusion requires effective modeling of complex networks. Traditional evaluation methods for static networks often fall short when applied to temporal networks. This paper introduces DGSP-GCN (Dynamic Graph Similarity Prediction based on Graph Convolutional Network), a deep learning-based framework that integrates graph convolutional networks with dynamic graph signal processing techniques to provide a unified solution for evaluating generative models and detecting anomalies in dynamic networks. DGSP-GCN assesses how well a generated network snapshot matches the expected temporal evolution, incorporating an attention mechanism to improve embedding quality and capture dynamic structural changes. The approach was tested on five real-world datasets: WikiMath, Chickenpox, PedalMe, MontevideoBus, and MetraLa. Results show that DGSP-GCN outperforms baseline methods, such as time series regression and random similarity assignment, achieving the lowest error rates (MSE of 0.0645, MAE of 0.1781, RMSE of 0.2507). These findings highlight DGSP-GCN's effectiveness in evaluating and detecting anomalies in dynamic networks, offering valuable insights for network evolution and anomaly detection research.

Abstract 3D object detection for self-driving cars has emerged as a critical challenge, largely due to the constraints of computing resources and the requirement for real-time processing. For evaluation, we utilize the KITTI benchmark dataset, which is widely used for self-driving cars and 3D object detection research. In this paper, an adaptive Weighted knowledge distillation approach is proposed to enhance detection accuracy while improving computational efficiency. Based on the performance of the teacher model, point clouds are divided into TPS (Teacher Performs Strongly) and TPW (Teacher Performs Weakly). The student model adapts its learning strategy dynamically: for TPS point clouds, it closely imitates the teacher by increasing the distillation weight, whereas for TPW point clouds, it prioritizes learning from raw data, reducing reliance on the teacher’s guidance. Additionally, a data pruning mechanism creates a smaller dataset from the KITTI benchmark based on the teacher model’s performance, while maintaining the TPS-TPW ratio. Experimental results indicate that the student model achieves comparable, and in some cases superior, performance to the teacher model. Specifically, it enhances recall by up to 1.5% and precision by up to 2.2% in complex scenarios. The student model is a lightweight network that learns from the teacher through knowledge distillation [28] and is solely used during execution. This design reduces execution time by nearly 50% while maintaining high detection accuracy. These findings emphasize the effectiveness of the proposed framework in real-time 3D object detection, making it well-suited for deployment in resource-constrained environments such as self-driving cars.

Abstract With the rapid expansion of the Internet and social networks, the volume of images available online has increased dramatically, making it challenging for users to find relevant content. To address this problem, we propose incorporating emotion analysis as a key factor in understanding user preferences, thereby creating a more personalized and effective image recommendation system. In this article, we examine two approaches to utilizing emotional features in image recommendation. The first approach integrates emotional features directly into the feature vector used for training the recommendation model. The second approach refines recommendations through emotion-based postprocessing, where emotional proximity between users and images is used to re-rank recommendations. This study emphasizes the value of emotion analysis in advancing the personalization and efficacy of social image recommendation systems. Experimental results indicate that both approaches significantly improve recommendation performance, achieving higher metrics such as Recall@k and Precision@k. These findings demonstrate that emotional analysis enhances personalization and effectiveness in social image recommendation systems.

Abstract Graph-structured data, common in real-world applications, captures entities (nodes) and their relationships (edges). While traditional methods integrate node content and neighborhood information to represent nodes in a latent space, random walks—despite being grounded in graph topology—suffer from limitations such as bias towards high-degree nodes, slow convergence, and difficulty in handling disconnected components. To address these issues, we introduce the "Feature-Based Random Walk on Graphs" (FARW), an advanced method that prioritizes node similarity in random walks. Unlike traditional approaches, FARW determines movement based on node features, enabling a more comprehensive analysis of complex networks. This feature-based approach improves the representation of heterogeneous graphs and enhances performance on a variety of tasks. Moreover, FARW demonstrates greater robustness when the graph structure changes. Experiments on three datasets—Cora, PubMed, and CiteSeer—show that FARW outperforms traditional structure-based random walks and the Node2Vec method, achieving accuracies of 87%, 83%, and 65%, respectively. These results suggest that incorporating node features during random walks improves the efficiency and accuracy of network analysis across diverse applications

Abstract Tokenization is a critical stage in text preprocessing and presents numerous challenges in languages like Persian, where there is no deterministic word boundary. These challenges include the identification of multi-function morphemes, separation of punctuation marks, omission of spaces between tokens, and handling extra spaces inside words.



Typically, the evaluation of tokenizers focuses on overall performance, and test data does not necessarily cover all challenging linguistic phenomena. As a result, strengths and weaknesses of tokenizers in addressing specific challenges are not independently assessed. This paper examines the challenges posed by the Persian script in detecting word boundaries and evaluates the performance of seven tokenizers in handling these issues. A test set of 4091 tokens across 483 sentences was prepared, with 1010 considered as challenging tokens. The tokenizers were evaluated using this dataset.



The results indicate varying performance among tokenizers when dealing with Persian orthography. Some tokenizers performed better in separating compound words, while others excelled in identifying and preserving zero-length joiners (half-space). A detailed comparison reveals that no tokenizer fully addresses all challenges, highlighting the need for improved algorithms and more sophisticated solutions for Persian word boundary detection.



By introducing a comprehensive benchmark and identifying the strengths and weaknesses of available tokenizers, this study paves the way for the development of better Persian language processing tools.

Abstract The stock market plays an imperative role in the entire financial market. The intricate and multifaceted nature of the stock market poses a challenge for investors seeking to establish a reliable and profitable trading approach. This paper aims to address this issue by leveraging two methodologies based on Deep Reinforcement Learning (DRL), namely Deep Q-Network (DQN) and Deep Deterministic Policy Gradient (DDPG), incorporating Convolutional Neural Network (CNN) and Gate Recurrent Unit (GRU) architectures, along with an attention mechanism to boost the decision making based on time-series stock data. This adaptation enables the model to focus on essential features and time periods within the stock data, leading to more successful and higher-quality trading choices. Following extensive experimentation and analysis, our proposed RLbased trading demonstrates improved accuracy and profitability compared to similar approaches. The proposed methodology strives to offer investors a dependable and lucrative trading strategy, ultimately leading to a more prosperous and efficient stock trading experience.

Abstract K-means clustering is one of the most prominent clustering methods that is used in many applications. By considering a widespread application of k-means clustering, redesign of this method in the context of high-performance computing has a considerable impact. In this paper, we consider scalability and utilize the available resources at a different level of parallelism. As a result, novel techniques are proposed for different hardware platforms, which are evaluated separately on uniformly random generated datasets and with different sizes. We change classic two-stage Lloyd’s formulation to a three stage that utilizes different techniques for each stage separately. Besides, we use an algebraic technique to reduce the amount of calculation and lay the foundation for consequent ideas. In CPUs, we propose a parallel architecture based on OpenMP and AVX2 instruction set. In GPUs, we utilize atomic operation and shared memory without considering GPU memory, and shared memory capabilities. Proposed method extends to multi-GPU. We merge these techniques and utilize MPI to scale it for multiple-node platforms.

Abstract Image captioning aims at automatically generating description of an image in natural language. This is a challenging problem in the field of artificial intelligence that has recently received significant attention in the computer vision and natural language processing. Among the existing approaches, visual retrieval based methods have been shown to be highly effective. These approaches search for similar images, then build a caption for the query image based on the captions of the retrieved images. In this study, we present a method for visual retrieval based image captioning, in which we use a multi criteria decision making algorithm to effectively combine several criteria with proportional impact weights to retrieve the most relevant caption for the query image. The main idea of the proposed approach is to design a mechanism to retrieve more semantically relevant captions with the query image and then selecting the most appropriate caption by imitation of the human act based on a weighted multi-criteria decision making algorithm. Experiments conducted on MS COCO benchmark dataset have shown that proposed method provides much more effective results compared to the state-of-the-art models.

Abstract In this paper, a neural approach based on Long Short-Term Memory (LSTM) neural networks is proposed to diagnose patients suffering from Parkinson’s Disease (PD). Considering the movement disorders caused by PD, the proposed method investigates the gait cycle pattern of subjects based on vertical Ground Reaction Force (vGRF) measured by 16 wearable sensors placed in subjects' shoes. In this study, it is shown that the temporal patterns of the gait cycle are different for healthy persons and patients. Therefore, by using a recurrent structure like LSTM, able to analyze the dynamic nature of the gait cycle, the proposed method extracts the temporal patterns to diagnose patients from healthy persons. To reduce the number of data dimensions, the sequences of corresponding sensors measuring vGRF in different feet are combined by subtraction. This method analyzes the temporal pattern of time series collected from different sensors, without extracting special features representing statistics of different parts of the gait cycle. Indeed, the method can extract temporal features based on learning, without using expert knowledge. Finally, the Accuracy and F1 Score of the model trained with all data is $99.87\%$, and $96.66\%$ respectively.