2026-08-17T08:58:57+03:30 http://jsdp.rcisp.ac.ir/browse.php?mag_id=61&slc_lang=fa&sid=1

61-1394 2026-08-17 10.1002

Signal and Data Processing JSDP 2538-4201 2538-421X 10.66224/jsdp 2025 22 2 Mahdi Sadeghi Moghadam mahdi.s.m.1366@gmail.com samad nejatian s.nejatian@gmail.com Hamid Parvin parvin@alumni.iust.ac.ir Karamullah Bagheri Fard k.bagherifad@gmail.com Seyed Hadi Yagoubian h.yaghoobian@gmail.com Optimization issues are often defined statically, assuming constant environmental conditions. However, in many real-world scenarios, problem environments are dynamic and continuously changing. Thus we need optimization algorithms that could solves those issues in dynamic environments as well. Dynamic optimization problems are change(s) that may occur through the time. Such environments are characterized by uncertainty, temporal changes, and structural complexities, which makes the optimization process a significant challenge. In addressing these challenges, evolutionary algorithms have emerged as one of the most effective approaches for solving dynamic optimization problems (DOPs). Among these algorithms, the Imperialist Competitive Algorithm (ICA), designed based on swarm intelligence and competition among imperialist countries, has garnered considerable attention due to its capability in solving static optimization issues. In this research, Imperialist Competitive Algorithms, inspired by the historical and political processes of colonization and assimilation, have been known as one of the efficient evolutionary algorithms. These algorithms face numerous challenges when dealing with dynamic problems , including reduced population diversity, performance degradation in conditions of rapid environmental changes, and limitations in optimal convergence. These cases indicate the need to develop improved and more adaptable versions of these algorithms. Using concepts such as memory, population clustering, and repulsion mechanisms, this algorithm has been able to maintain population diversity at all stages while increasing the speed of convergence in the face of environmental changes. The key feature of the proposed algorithm is the use of memory to store previous optimal solutions, a clustering mechanism to manage population diversity, and repulsion to prevent unnecessary accumulation of solutions in specific regions. Nevertheless, ICA exhibits poor performance in dynamic environments because it lacks mechanisms to maintain diversity, quick adaptation to environmental changes, and new optima track. This study presents an improved version of ICA aimed at overcoming these limitations. The proposed algorithm incorporates a combination of a memory mechanism and a clustering strategy to enhance its adaptability to environmental changes and preserve diversity within the population. The memory mechanism stores information about previous optima and utilizes it under appropriate conditions to accelerate the optimization process.for clustering method is used for clustering. Clustering in the proposed method ensures that diversity is maintained for the population during the execution of the algorithm. In this study, our goal is to solve problems that change the environment in a global way. That means, the fitness of all points in the environment changes. By testing just one point in the environment and comparing the fitness obtained with its previously stored value, we can detect a change in the environment. On the other hand, the clustering strategy, particularly the k-means technique, to maintain maintain population diversity and prevents the convergence of solutions to specific regions. Together, these two components create a balance between exploration and exploitation, thereby improving the algorithm's performance in dynamic environments. To evaluate the performance of the proposed algorithm, the Moving Peaks Benchmark (MPB) was used as a standard metric. Due to its capability to simulate complex and diverse changes in dynamic environments—particularly in Branke's second scenario—MPB is one of the most recognized tools for assessing the performance of dynamic optimization algorithms. The proposed algorithm was evaluated alongside advanced algorithms such as FTmPSO (TMO), RAmQSO-s4, RmNAFSA-s4, TFTmPSO, RFTmPSO, mQSO10 (5+5q), FMSO, CellularPSO, Multi-SwarmPSO, mCPSO, AmQSO*, FTMPSO, almPSO, and CDEPSA. Experimental results demonstrated that the proposed algorithm outperformed other methods in areas such as convergence speed, adaptability to environmental changes, and population diversity preservation. A key feature of the proposed algorithm is its ability to retain identified optima even after environmental changes. Additionally, the use of the k-means clustering technique has ensured that the algorithm effectively avoids excessive focus on specific regions and maintains population diversity while facing complex environmental changes. Another advantage of this algorithm is its scalability in handling dynamic optimization issues with high dimensions and complexities. These findings indicate that the proposed algorithm is not only effective in laboratory settings but also suitable for real-world applications with fast and dynamic changes. Dynamic Optimization Dynamic Environments Memory Imperialist Competition Algorithm Clustering Moving Peaks Benchmark. 2025 9 01 3 30 http://jsdp.rcisp.ac.ir/article-1-1394-en.pdf 10.61882/jsdp.22.2.3

61-1203 2026-08-17 10.1002

Signal and Data Processing JSDP 2538-4201 2538-421X 10.66224/jsdp 2025 22 2 Zahra Shaeeri shaeiri.zahra@gmail.com Javad Kazemitabar j.kazemitabar@nit.ac.ir Soroush Haghverdi haghverdi@ifb.ir Stock market fraud, particularly front-running, is a deceptive practice in which traders exploit prior knowledge of significant orders placed by others to profit from stock price movements. Front-running is considered illegal because it involves using confidential or non-public information to manipulate the market for personal gain. This paper tries ti propose a novel, unsupervised, and real-time anomaly detection method based on behavioral analysis, specifically designed to identify front-running fraud within stock market transactions. The method focuses on building individual behavioral profiles for each trader, capturing their specific traits and patterns in stock buying and selling. These profiles serve as baselines for what is considered as 'normal' trading behavior for each trader. To detect anomalies, we introduce a statistical framework where the risk of each transaction will be calculated by evaluating the deviation from the expected behavior based on the trader's historical actions. This deviation is a measure of how unusual the current transaction is in comparison to the trader’s typical actions. The risk calculation involves the use of the log-likelihood ratio, a concept derived from detection theory, which compares the likelihood of a transaction being normal or fraudulent. The conditional probability of a transaction being either fraudulent or non-fraudulent is computed, and the ratio of these probabilities has been taken on a logarithmic scale to define the transaction risk. This risk metric is then utilized to flag potentially suspicious behavior for further investigations. Bayesian probability theory underpins the model, specifically employing Bayes' rule to update the likelihood of fraud as more data will be accumulated over time. The model assumes the independence of risk components, which simplifies the complexity of the system and improves computational efficiency. Despite the potential limitation of assuming independence, empirical studies have shown that this assumption often yields reliable results for detecting anomalous behavior, making the approach both practical and effective. Behavioral profiling plays a key role in this method. By observing the individual’s trading history—such as the frequency, timing, and amounts of trades—the system learns a trader’s typical behavior. This behavioral information is critical because it accounts for the natural variance in a trader's actions over time, allowing the model to distinguish between normal fluctuations and abnormal activities that might indicate fraud. Key behavioral indicators include the timing of trades, the volume of trades, the frequency of transactions with specific counterparties, and the trader’s overall market engagement. Traders whose actions significantly deviate from their established patterns—such as purchasing large quantities of stocks at unusual times or interacting with the same trader excessively—are flagged as high-risk. The simulation section of the paper uses 16 months of stock market transaction data, where features such as transaction amounts, time of trade, urgency, and consistency in trading with particular traders are analyzed to calculate the risk profile. The system ranks traders based on the risk scores of their transactions, enabling the detection of front-running activities in near real-time. The results from the simulation indicate that the proposed method is highly effective in identifying front-running fraud. The use of behavioral profiling ensures that the system is adaptive to individual trading patterns, making it resistant to the evolving nature of fraud in financial markets. The methodology also provides a significant advantage over traditional rule-based systems, which often struggle to adapt to new fraud techniques as they emerge. Furthermore, this approach can be applied in live trading environments, making it a practical tool for regulatory bodies and market surveillance. This paper contributes to the growing field of financial fraud detection by introducing an innovative approach that combines behavioral analysis with advanced statistical techniques. The findings underline the importance of real-time monitoring and adaptive fraud detection systems in maintaining market integrity. In the simulation section, stock market data of 16 months is used. Features related to amounts, hours, urgency, and trading with one trader in buying/selling have been used to obtain the ranking. Results show that the proposed method is effective in detecting front running cases Stock market fraud detection behavioral profiling data analytics Front running log-likelihood Bayes Rule anomaly detection. 2025 9 01 31 42 http://jsdp.rcisp.ac.ir/article-1-1203-en.pdf 10.61882/jsdp.22.2.31

61-1378 2026-08-17 10.1002

Signal and Data Processing JSDP 2538-4201 2538-421X 10.66224/jsdp 2025 22 2 ZAHRA sarhadi z_sarhadi72@yahoo.com mehdi khazaiepoor mkhazaiepoor@gmail.com “The Internet of Things” is an extensive network of intelligent objects that has a large number of objects connected to the Internet. One of the applications of the IOT network is in smart cities. In smart cities, all parts of the city, such as the transportation system, electricity network, health network, etc., are interconnected with IOT support. One of the critical challenges of the IOT network is the occurrence of attacks against this network, which causes the network services to be disrupted. Intrusion detection systems are used to detect attacks on the IOT. The role of an IoT network intrusion detection system is to analyze the network traffic, detect abnormal traffic, and send necessary warning to the firewall. One of the methods of detecting attacks on the IOT and smart cities is to use machine learning methods such as support vector machines(SVM). One method to reduce the error of the support vector machine in detecting attacks on the IOT network and the smart city is to use feature selection methods and optimize its parameters. By selecting the feature and optimizing the parameters of the support vector machine, the attack detection error will be reduced. In this article, an intrusion detection method with an artificial neural network and a swordfish optimization algorithm is presented to detect attacks on the smart city. The proposed method includes three different phases: data set balancing with game theory and GAN network, feature selection with Sailfish Optimizer algorithm, and optimization of SVM parameters with Archimedes optimization algorithm (AOA) algorithm. The role of a multilayer neural network in the proposed method of evaluating feature vectors and the role of the support vector machine is to classify network traffic into two categories: attack and normal. The evaluation and tests performed in MATLAB software and on the NSL-KDD data set show that the accuracy, sensitivity, and precision of the proposed method are 99.12%, 98.92%, and 98.96%, respectively, and the support vector machine with Gaussian kernel seems to be more accurate. The results of the experiments showed that the proposed method is more accurate than meta-heuristic algorithms, such as gray wolf optimization and genetic algorithms in detecting attacks on the smart city Internet of Things Smart Cities Intrusion Detection System Machine Learning Sailfish Optimizer Algorithm Feature Selection 2025 9 01 43 64 http://jsdp.rcisp.ac.ir/article-1-1378-en.pdf 10.61882/jsdp.22.2.43

61-1451 2026-08-17 10.1002

Signal and Data Processing JSDP 2538-4201 2538-421X 10.66224/jsdp 2025 22 2 Fatemeh Amoli amoli.fatemeh.1996@gmail.com Mostafa Bastam bastam@umz.ac.ir Ehsan Ataei ataie@umz.ac.ir With the rapid rise of cybersecurity threats and the increasing complexity of digital security, event log data serves as a critical source for identifying and analyzing cyberattacks and threats. This data provide key insights into system activities, essential for detecting unauthorized intrusions, analyzing suspicious behaviors, and conducting security investigations. However, any alteration or tampering with the data can disrupt the analysis and detection processes, leading to incorrect security decisions. Blockchain technology, with its unique features such as decentralization, immutability, and transparency, has been recognized as a reliable and secure platform for storing and protecting data. This technology enables the storage of data hashes in a way that any changes can be easily detected. However, directly storing the vast volume of event log data on the blockchain faces challenges such as high costs and storage space limitations. In this research, an innovative model has been presented to automate the assurance of event log data integrity and confidentiality using the public Ethereum blockchain and smart contracts. Instead of storing event log data directly, only their hashes have been saved on the blockchain. This approach not only reduces storage costs but also ensures data confidentiality. The automated data integrity assurance process in this model occurs in two stages: Stage One: Event log data hashes have been periodically stored on the blockchain and compared with previous hashes. Stage Two: Over longer intervals, all stored hashes have been reviewed and validated to prevent any potential tampering. In this study, the costs associated with implementing this model on the Ethereum Sepolia test network had been precisely calculated. The analysis indicates that operational costs and computational overhead have been optimized across different time intervals, demonstrating the model's feasibility for large-scale deployment. Ultimately, this research tries to introduce a novel and practical model, taking a significant step toward automating the assurance of event log data integrity and confidentiality, providing a reliable solution for real-world applications. Log management Data integrity Blockchain Ethereum Smart Contract. 2025 9 01 65 78 http://jsdp.rcisp.ac.ir/article-1-1451-en.pdf 10.61882/jsdp.22.2.65

61-1454 2026-08-17 10.1002

Signal and Data Processing JSDP 2538-4201 2538-421X 10.66224/jsdp 2025 22 2 Babak Azarnavid babakazarnavid@ubonab.ac.ir Mohsen Abdolhosseinzadeh mohsen.ab@ubonab.ac.ir Hojjat Emami emami@ubonab.ac.ir Liver diseases are among the leading causes of mortality worldwide, deeply influencing individuals' lives, often at younger ages when they are in the prime of their personal and professional lives. The insidious nature of these diseases lies in their early initial symptoms, which frequently goes unnoticed until the condition has progressed to an advanced stage. This delay in diagnosis not only diminishes the chances of successful treatment but also places an immense emotional and financial burden on patients as well as families. Early detection is therefore critical, as it can significantly alter the course of the disease, improving survival rates and quality of life. However, traditional diagnostic methods often fall short in terms of speed, accuracy, and accessibility, particularly in resource-limited settings. This underscores the urgent need for innovative approaches to liver disease detection and its management. Machine learning (ML) has been emerged as a powerful tool in this regard, offering the potential to revolutionize how we diagnose and predict liver diseases. By leveraging vast datasets—ranging from clinical records and laboratory results to imaging data—ML algorithms can uncover complex patterns and correlations that may elude human experts. These insights can lead to earlier and more accurate diagnoses, enabling timely interventions that can save lives. Among the various ML approaches, stacked machine learning (SML) models stand out for their ability to combine the strengths of multiple algorithms, mitigating the limitations of individual models and enhancing overall performance. This research focuses on developing and evaluating an SML model specifically designed for the accurate diagnosis, classification, and prediction of liver diseases, with the goal of addressing some of the most pressing challenges in this field. The proposed SML model employs a sophisticated two-layer architecture to tackle common issues such as overfitting and improving prediction accuracy. In the first layer, the model integrates four robust base learner algorithms: Extremely Randomized Trees (ET), Decision Tree (DT), Random Forest (RF), and Extreme Gradient Boosting (XGB). Each of these algorithms contributes unique strengths, such as handling high-dimensional data, capturing non-linear relationships, and reducing variance. The predictions generated by these base learners are then fed into the second layer, where a Logistic Regression (LR) algorithm synthesizes the outputs to produce the final prediction. This layered approach ensures that the model benefits from the collective intelligence of multiple algorithms, resulting in more reliable and precise outcomes. To further optimize performance, the Grid Search (GS) algorithm was employed to fine-tune the parameters of the learning algorithms, ensuring that the model operates at its full potential. This study employs dataset from the University of California, Irvine (UCI) Machine Learning Repository. A sample size of 615 instances has been utilized to implement the proposed methodologies, with a stratified division of 70% for training and 30% allocated for testing purposes. The results of this research seems to be highly promising. Evaluation based on 5-fold cross-validation demonstrates that the proposed SML model outperforms existing methods, achieving an impressive 0.9940 accuracy and a 0.9880 F1-score on the test data. These metrics not only highlight the model's exceptional predictive capabilities but also underscore its potential to serve as a valuable tool for clinicians in real-world settings. By providing accurate and timely diagnoses, the SML model can help reduce the mortality and morbidity associated with liver diseases, offering hope to patients and their families. Beyond the technical achievements, the human impact of this research cannot be overstated. For patients, the SML model represents a lifeline—a chance to detect liver diseases early, when treatment seems most effective, and to avoid the devastating consequences of late-stage diagnoses. For healthcare providers, it offers a reliable and efficient diagnostic tool that can enhance decision-making and improve patient outcomes. Also, for society as a whole, it signifies a step forward in the fight against a disease that disproportionately affects vulnerable populations, including those in underserved regions where access to advanced medical care is limited. In essence, this research is not just about developing a sophisticated algorithm; it is also about harnessing the power of machine learning to make a tangible difference in people's lives. By bridging the gap between cutting-edge technology and human care, the proposed SML model embodies the potential of computer science to address some of the most critical health challenges of our time. It is a testament to the transformative power of innovation, compassion, and collaboration in the pursuit of better health for all. Liver diseases Early Diagnosis Machine Learning Cumulative Machine Learning Model Cross-Validation 2025 9 01 79 96 http://jsdp.rcisp.ac.ir/article-1-1454-en.pdf 10.61882/jsdp.22.2.79

61-1376 2026-08-17 10.1002

Signal and Data Processing JSDP 2538-4201 2538-421X 10.66224/jsdp 2025 22 2 fatmhjlaly128@gmail.com Ateke Goshvarpour ak_goshvarpour@imamreza.ac.ir Background: Electroencephalography (EEG) is a cornerstone in cognitive neuroscience, providing critical insights into the neural mechanisms underlying skill acquisition. Despite significant advancements in signal processing techniques, extracting meaningful patterns from EEG data —especially in the context of dynamic neural shifts during learning—remains a persistent challenge. Traditional analytical approaches often fail to account for the nonlinear temporal dynamics inherent in learning processes, which limits their ability to decode subtle neural reorganizations. This study addresses this gap by proposing an innovative computational framework based on Rhodonea curves—sinusoidal patterns resembling flower petals—to analyze EEG signals during the acquisition of a complex motor skill: touch-typing by the Colemak keyboard layout. Objective and Innovation: The study aims to develop and validate a computationally efficient algorithm for classifying EEG data across distinct stages of skill learning. Central to this approach is the introduction of asymmetry indices derived from Rhodonea curves, which quantify nonlinear features of brain activity. This work represents the first application of Rhodonea-based analysis in EEG signal processing, providing a geometrically intuitive and computationally lightweight alternative to conventional nonlinear methods, such as entropy or fractal dimension analysis. Methodology: The dataset, available on IEEEDataPort, consisted of EEG recordings from 10 participants (6 females and 4 males), focusing on 9 channels (F3, Fz, F4, C3, Cz, C4, P3, POz, P4) collected during 12 typing sessions. Data from sessions 4, 8, and 11—representing the early, intermediate, and advanced learning phases—were analyzed, with each session repeated five times to capture intra-session variability. For the first time, a Rhodonea curve-based method has been introduced for signal analysis, featuring a structure resembling a flower with an adjustable number of petals. The Rhodonea model was parameterized with one to ten petals, and three new indices based on asymmetry in the Rhodonea curve were computed to characterize spatiotemporal variations in EEG signals. A Support Vector Machine (SVM) utilizing a one-vs-all strategy was employed to classify 15 classes (5 repetitions × 3 sessions). Channel-specific optimizations and petal-count analyses were conducted to identify discriminative brain regions and optimal model configurations. Key Findings: The analysis revealed robust classification performance, with two-class classification achieving accuracies ranging from 79.3% to 93.3%. Optimal results were observed in channels F3, Fz, C3, C4, and POz using a 4-petal Rhodonea configuration. In the three-session classification, the highest accuracy was recorded for the advanced learning phase (Session 11: 92%), followed by the early phase (Session 4: 90%) and the intermediate phase (Session 8: 72.6%). The lower accuracy in Session 8 suggests a transitional neural state marked by unstable skill consolidation, where neither novice nor expert patterns dominate. Neuroanatomically, the frontal (F3, Fz), central (C3, C4), and parieto-occipital (POz) regions demonstrated heightened discriminative power, consistent with prior studies implicating these areas in cognitive control, motor planning, and visuospatial integration during learning. Session-specific activation patterns indicated early-phase prefrontal engagement for attention allocation and advanced-phase parietal consolidation for skill automatization. Comparative Analysis: This study diverges from prior work by integrating geometric asymmetry metrics—rather than spectral or entropy-based features—to model learning-induced neural plasticity. The computational efficiency and interpretability of Rhodonea-based features (e.g., petal-count visualization) offer distinct advantages for real-time brain-computer interface (BCI) applications. Notably, the intermediate phase’s lower accuracy (72.6%) highlights the methodological challenge of decoding transitional neural states, a limitation underrepresented in earlier literature. Limitations and Future Directions: This research had limitations that should be considered in future studies. First, the small sample size (N=10) and fixed signal length (1280 samples) may limit generalizability; future work should incorporate larger datasets and variable-length signal analysis. Second, although the non-linear features presented are computationally simple and low-cost, using other complex features might enhance the model's performance. Third, while SVM demonstrated efficacy, comparative studies with deep learning models (e.g., CNNs, LSTMs) could further validate the method’s robustness. Fourth, physiological validation via multimodal neuroimaging (e.g., fMRI/fNIRS) is needed to spatially localize the observed dynamics. Finally, statistical refinements—such as ANOVA or t-tests for feature selection—could enhance model rigor and mitigate overfitting risks. Conclusion: This research pioneers the application of Rhodonea curves in EEG analysis, establishing a novel framework for decoding the neural correlation of skill learning. The high classification accuracies and neuroanatomically consistent results underscore the method’s potential for both academic research and applied domains, including adaptive learning systems and neurorehabilitation. Future efforts should prioritize large-scale validation and integration with multimodal neuroimaging to advance our understanding of learning-related brain plasticity and refine real-world applications. Electroencephalography Learning Rhodonea curve Signal processing Classification Asymmetry 2025 9 01 97 108 http://jsdp.rcisp.ac.ir/article-1-1376-en.pdf 10.61882/jsdp.22.2.97

61-1252 2026-08-17 10.1002

Signal and Data Processing JSDP 2538-4201 2538-421X 10.66224/jsdp 2025 22 2 Saeede Anbaee Farimani anbaee@mshdiau.ac.ir Raheleh Ghouchannezhad noor nia rghoochannejad@yahoo.com MAJID VAFAEI JAHAN vafaeijahan@mshdiau.ac.ir The onset of social media venues, online news media, and digital content allowed a vast volume of text and time series data to be generated which plays significant role in investors' decision-making and financial market volatility. Data extracted from these platforms provide information on public sentiments, immediate reactions to news, and informal analyses, which, if processed appropriately, can be very useful indicators in forecasting financial market trends. Billions of dollars are invested and lost, depending on correct forecasting. However, advances in deep learning, especially in large language models (LLMs) and novel time series analysis algorithms, have opened new windows to processing and analyzing this complex data. The advanced language models identify hidden patterns and nonlinear dependencies, always taking into account the context and semantic details of the text between news, market sentiments, and price fluctuations, as well as utilizing them via intelligent market analysis systems. This review analyzes the existing research trends on the relationship of text data available on websites and social networks with the behavior of financial markets, having reviewed more than 200 scientific papers published between 2006 and 2024 in a systematic manner. This study focuses on identifying advanced methods within text representation, sentiment analysis, predictive modeling, and language model applications for analyzing real-time and unstructured data. More than one information source has to be taken into consideration: (Twitter, news agencies, blogs, and specialized forums) from a perspective of credibility, data structure, and influence-on market decisions. Given the complexity of financial markets, such as stocks and forex, there is an ever-increasing demand for hybrid models capable of carrying out analyses across time-series and text data simultaneously. This paper aims to analyze the current research accomplishments, identify gaps in the research, and ultimately put forward future directions for the fields of text mining, AI, and deep learning. These directions can open up the path for the next generation of real-time and adaptive recommender, predictor, and correlation analyzer systems in the financial markets. Large Language Models Text Mining Sentiment Analysis Financial Market Prediction News Social Media 2025 9 01 109 126 http://jsdp.rcisp.ac.ir/article-1-1252-en.pdf 10.61882/jsdp.22.2.109

61-1433 2026-08-17 10.1002

Signal and Data Processing JSDP 2538-4201 2538-421X 10.66224/jsdp 2025 22 2 Mohsen Nasiri program.nasiri97@gmail.com Negin Daneshpour ndaneshpour@sru.ac.ir The problem of classification can be divided into two categories: single-label and multi-label. Single-label classification consists of binary and multi-class classification. In binary classification, the task is to predict one in two possible classes, such as distinguishing between spam and non-spam emails. In multi-class classification, the goal is to classify instances into more than two classes, such as identifying different species of flowers based on petal measurements. In contrast to single-label classification, multi-label classification is more complex because each instance could belong to multiple categories simultaneously. In multi-label learning, instead of assigning a single label to each instance, a set of labels is assigned. This means that each sample may have zero, one, or more than one associated label. For example, in a text classification task, a news article about technology and business might be labeled as both "Technology" and "Business". To handle multi-label classification, several approaches have been developed. One of the simplest methods is Binary Relevance (BR), which transforms the multi-label problem into multiple independent binary classification tasks—one for each label. Although this approach is easy to implement, it treats each label independently and ignores possible relationships among them. However, in real-world applications, labels are often correlated; for instance, in medical diagnosis, certain diseases frequently appear together. In another approach, Label Powerset (LP), considers label dependencies by treating each unique combination of labels as a separate class. While this method captures relationships between labels, it suffers from scalability issues while dealing with a large number of labels, as the number of possible label combinations increases exponentially. To address these challenges, the proposed method incorporates k-means constraint clustering to group both labels and features prior to applying classification. In the first step, clustering is performed to group similar labels together, ensuring that label correlations are preserved. This also helps to mitigate the issue of imbalanced classification, where certain labels may be underrepresented in the dataset. Once the labels are being clustered, a separate multi-layer neural network would be assigned to each cluster. Instead of using a single large neural network for all labels, multiple smaller networks would be trained for different label clusters. This approach enhances learning efficiency and improves accuracy by focusing on relevant label groups. However, using multiple classifiers increases computational costs and training time. To mitigate this issue, a scatter-add dimension reduction technique is applied. Using scatter-add, attributes are efficiently assigned to the input of each neural network, ensuring that each classifier receives only the relevant feature subset. Each neural network then predicts labels within its designated cluster. Eventually, the predictions from all classifiers are combined to generate the final multi-label output for each instance. To evaluate the effectiveness of the proposed method, experiments were conducted on various text datasets. The results were compared with traditional multi-label classification methods, including Binary Relevance and Label Powerset. The evaluation has been based on several performance metrics, such as accuracy, precision, and hamming-loss. The results demonstrated that the proposed approach achieved superior performance across multiple datasets, ranking first in several evaluation criteria. Notably, it outperformed existing methods by a margin of approximately 1% in accuracy. These findings suggest that clustering-based multi-label classification using k-means constraint clustering and multi-layer neural networks is a promising approach. By leveraging label correlations and reducing dimensionality, the proposed method effectively improves classification performance while addressing issues such as label imbalance and computational inefficiency. Future research may further explore optimization techniques to reduce training time while maintaining high accuracy. Classification Multi-Label Classification Clustering Neural Networks 2025 9 01 127 138 http://jsdp.rcisp.ac.ir/article-1-1433-en.pdf 10.61882/jsdp.22.2.127