Journal of Digital Technology and Computer Science
Digital Technology and Socio-Technical Innovation, including the design, development, implementation, and evaluation of digital solutions, platforms, applications, and infrastructures that support modern socio-technical systems, digital transformation, and technology-enabled services. Computer Systems, Software, and Networking, encompassing distributed systems, computer networks, network architectures, communication protocols, network performance, next-generation connectivity, software systems, and integrated computing environments. Artificial Intelligence, Machine Learning, and Intelligent Systems, covering intelligent systems, machine learning algorithms, deep learning, natural language processing, expert systems, knowledge-based systems, computational intelligence, and applied AI across scientific, industrial, and societal domains. Decision Support, Fuzzy, and Evolutionary Systems, including decision support systems, fuzzy logic, fuzzy control, evolutionary computing, optimization algorithms, swarm intelligence, hybrid intelligent methods, and data-driven decision models. Image, Audio, and Multimedia Processing, including computer vision, image processing, sound and speech processing, multimedia analysis, signal processing, pattern recognition, and audio-visual computing applications. Information Security and Cybersecurity, focusing on information security, system security, network security, cybersecurity governance, cryptography, privacy protection, secure software engineering, threat detection, intrusion prevention, digital forensics, and cyber risk management. Cyber Crime and Digital Investigation, including cybercrime detection and analysis, cyber law and policy in digital environments, forensic investigation, online fraud, identity theft, malicious activity analysis, and digital evidence management. Social Network and Digital Security, covering security and trust in social media and online platforms, digital identity, misinformation and disinformation detection, privacy in social networks, human factors in cybersecurity, and safe digital interaction ecosystems. Operating Systems, Computer Architecture, and Embedded Computing, including operating systems, processor and memory architecture, virtualization, system-level optimization, embedded systems, real-time computing, firmware, and performance engineering. Cloud, Edge, and Ubiquitous Computing, covering cloud platforms, fog and edge computing, distributed intelligence, service orchestration, scalable infrastructures, reliability, resource management, and pervasive computing environments. Internet of Things (IoT), Sensor Networks, and Cyber-Physical Systems, including smart devices, wireless sensor networks, industrial IoT, IoT platforms, connected environments, cyber-physical systems, and real-world deployment challenges in intelligent sensing and control. Big Data, Analytics, and Data-Driven Computing, encompassing data engineering, data mining, large-scale data processing, predictive analytics, visual analytics, business intelligence, and advanced computational methods for complex datasets. Wearable Devices and Smart Sensing Technologies, including wearable computing, body-area networks, health and activity monitoring systems, smart textiles, mobile sensing, and human-centered intelligent devices. Embedded Robotics and Microcontroller Systems, including robotic systems, embedded robotics, autonomous control, low-level hardware-software integration, microcontroller-based applications, robotic sensing, and intelligent actuation systems. Micro and Nano Technology, including microelectronics, nanoelectronics, microsystems, nanosystems, MEMS/NEMS-related applications, miniaturized intelligent devices, and sensor-oriented micro/nano technological innovations. Renewable Energy and Intelligent Energy Systems, including digital technologies for renewable energy, smart energy monitoring, intelligent control systems for energy efficiency, IoT-enabled energy systems, sustainable computing, and computational methods for energy optimization. Software Engineering and Information Systems, including software design, software quality assurance, software testing, requirements engineering, enterprise systems, information systems development, human-centered software solutions, and digital service integration. Robotics, Automation, and Autonomous Systems, covering intelligent robotics, automation systems, control engineering, autonomous agents, robotic perception, human-robot interaction, and smart manufacturing applications. Human-Computer Interaction and Digital Experience, including user interface design, user experience, usability evaluation, interactive systems, accessibility, persuasive technologies, and digital behavior in technology-mediated environments. Green Computing and Sustainable Digital Systems, including energy-efficient computing, sustainable software and hardware design, green AI, carbon-aware digital infrastructures, smart resource management, and digital technologies for environmental sustainability.
Articles
63 Documents
<![CDATA[Application of Fuzzy Multi Criteria Decision Making (FMCDM) and TOPSIS in Selecting Prospective Employees ]]>
<![CDATA[Lailatul Badria]]>;
<![CDATA[Sriani]]>
<![CDATA[ Journal of Digital Technology and Computer Science Vol. 3 No. 3 (2026): August 2026 ]]>
Publisher : <![CDATA[Academic Bright Collaboration ]]>
Show Abstract
|
Download Original
|
Original Source
|
Check in Google Scholar
|
DOI: 10.66053/dtcs.v3i3.1132
<![CDATA[Purpose – The primary objective of this research is to design and build a web-based Decision Support System (DSS) to evaluate and hire prospective candidates at Perumda Tirtanadi. To achieve this, a hybrid methodology combining Fuzzy Multi-Criteria Decision Making (FMCDM) and the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) is implemented. The proposed system is intended to reduce subjectivity and improve the effectiveness of the employee selection process. Methods – In this framework, FMCDM is utilized to translate subjective, linguistic evaluations into precise crisp values through a defuzzification process, which subsequently establishes the relative weights of the evaluation criteria. Afterward, these weights are integrated into the TOPSIS algorithm to calculate the final rankings of the applicants across nine specific benchmarks, namely GPA, Basic Competency Test (TKD), psychological test score, attitude, communication skills, work experience, politeness, field expertise, and personal development plan. The system was implemented as a web-based application using PHP and MySQL. Findings – The developed system successfully performed automatic FMCDM weighting and TOPSIS ranking. The ranking results were consistent with manual calculations, indicating that the proposed approach was correctly implemented and capable of producing accurate candidate rankings. Research implications – This study used recruitment data from a single organization, namely Perumda Tirtanadi Medan. Consequently, the applicability of the outcomes is restricted to the specific criteria and administrative environment of this particular utility company and may not be directly generalizable to other organizations. Originality – This study integrates FMCDM for criteria weighting and TOPSIS for candidate ranking into a web-based decision support system, providing a practical and objective approach to employee selection at Perumda Tirtanadi.]]>
<![CDATA[Sistem Penyiraman Otomatis Bibit Kelapa Sawit Berbasis IoT Menggunakan Metode Fuzzy Sugeno Berdasarkan Kelembapan Tanah Dan Suhu ]]>
<![CDATA[Adinda Puspita]]>;
<![CDATA[M. Fakhriza]]>
<![CDATA[ Journal of Digital Technology and Computer Science Vol. 3 No. 3 (2026): August 2026 ]]>
Publisher : <![CDATA[Academic Bright Collaboration ]]>
Show Abstract
|
Download Original
|
Original Source
|
Check in Google Scholar
|
DOI: 10.66053/dtcs.v3i3.1141
<![CDATA[Purpose – This study aims to design and implement an IoT-based automatic irrigation system for oil palm seedlings capable of determining irrigation duration based on soil moisture and ambient temperature conditions. The research addresses the limitations of manual irrigation, which may cause water deficiency or excess in the growing medium. Methods – The study employed experimental and engineering methods using an ESP32 DevKit V1 microcontroller as the control center. The system utilized an FC-28 soil moisture sensor and a DHT22 temperature sensor. Sensor data were processed using the Sugeno Fuzzy Logic method with a nine-rule base to determine irrigation duration. A relay controlled a 5V DC water pump, while Telegram was used for system monitoring and control. Findings – The system successfully generated irrigation decisions with predefined singleton outputs of 0, 5, 10, and 15 seconds and correctly executed all implemented Telegram monitoring and control commands Research implications – The system is limited by the use of the FC-28 resistive soil moisture sensor, which is susceptible to corrosion and requires periodic recalibration, while remote monitoring depends on network availability. Originality – This study differentiates itself from previous work by implementing an explicitly specified ESP32 DevKit V1, a DHT22 temperature sensor, and Telegram-based monitoring and control, whereas previous research employed a DHT11 sensor and website-based monitoring without explicitly identifying the microcontroller while using the same fuzzy input parameters and Sugeno Fuzzy Logic method.]]>
<![CDATA[Decision Tree Classification of Fear of Missing Out Levels in Generation Z Technology Product Purchases ]]>
<![CDATA[Kalfida Eka Wati Siregar]]>;
<![CDATA[Aidil Halim Lubis]]>
<![CDATA[ Journal of Digital Technology and Computer Science Vol. 3 No. 3 (2026): August 2026 ]]>
Publisher : <![CDATA[Academic Bright Collaboration ]]>
Show Abstract
|
Download Original
|
Original Source
|
Check in Google Scholar
|
DOI: 10.66053/dtcs.v3i3.1148
<![CDATA[Purpose – This study used a Decision Tree to classify score-defined Fear of Missing Out (FOMO) categories associated with Generation Z technology-product purchasing, identify discriminative questionnaire items, and evaluate the stability and interpretability of the classification structure. Methods – The dataset contained 500 respondents and 13 self-developed four-point Likert items. Corrected item-total correlations, Cronbach’s alpha, and a preliminary one-factor exploratory factor analysis were examined. A stratified 80:20 split was used for an unpruned Decision Tree and a cost-complexity-pruned sensitivity model selected by five-fold training-only cross-validation. A deterministic score rule, root-split information gain, model-wide importance, and 100 repeated stratified splits were also evaluated. Findings – The unpruned Decision Tree classified 92 of 100 holdout observations correctly (92.00% accuracy; 92.35% weighted F1; 86.64% macro F1). Training-only pruning reduced the tree from 49 to 35 nodes and produced 93.00% accuracy and an 88.99% macro F1. X1 had the highest primary root-split information gain (0.6805) and ranked first in 74 of 100 repeated splits, whereas X5 ranked first in 26. Repeated-split accuracy averaged 94.13%, but balanced accuracy and macro F1 averaged 88.52% and 88.89%, respectively. The deterministic score rule achieved 100%. Research implications – The Low, Moderate, and High categories were equal-width operational score intervals and have not been externally validated. Preliminary factor analysis supported one broad common factor, but the self-developed instrument still requires independent content, construct, and criterion validation. The target was derived from the same items used as predictors, minority-class performance varied across splits, and the available records did not document ethics approval, guardian consent, or minor assent for participants younger than 18. The findings therefore describe score-category discrimination rather than causal purchasing behaviour or validated psychological severity. Originality – This study integrates interpretable Decision Tree classification, training-only pruning, psychometric diagnostics, deterministic-rule comparison, multiple feature-importance perspectives, and repeated-split stability analysis for technology-purchasing FOMO patterns.]]>
<![CDATA[Random Forest Algorithm with SMOTE Technique for Classifying the Mental Health of Fresh Graduates Facing Career Competition ]]>
<![CDATA[Wily Supi Ramadani]]>;
<![CDATA[Sriani]]>
<![CDATA[ Journal of Digital Technology and Computer Science Vol. 3 No. 3 (2026): August 2026 ]]>
Publisher : <![CDATA[Academic Bright Collaboration ]]>
Show Abstract
|
Download Original
|
Original Source
|
Check in Google Scholar
|
DOI: 10.66053/dtcs.v3i3.1149
<![CDATA[Purpose – This study evaluated Random Forest with the Synthetic Minority Oversampling Technique (SMOTE) for classifying questionnaire-derived career-related psychological categories among fresh graduates and descriptively compared performance before and after class balancing. Methods – Data were collected from 250 fresh graduates using a self-developed 13-item, four-point Likert questionnaire covering seven operational dimensions. A stratified 80:20 holdout split was used for the primary comparison, while 5 × 5 repeated stratified cross-validation examined stability. Standard SMOTE and random oversampling were applied only to training data, and SMOTE-generated profiles were audited in both stored-integer and continuous-interpolation representations. Findings – On the selected holdout split, Random Forest without SMOTE achieved 96.00% accuracy and a 95.93% weighted F1-score, whereas the SMOTE model reached 100.00% for both measures by correcting two Moderate-to-Poor errors. Random oversampling achieved 98.00% holdout accuracy. Across 25 repeated validation folds, mean accuracy was 96.08% without SMOTE, 95.76% with SMOTE, and 95.92% with random oversampling; the baseline–SMOTE accuracy difference was not statistically significant (Wilcoxon p = 0.412). Research implications – This study combines a controlled comparison of baseline Random Forest, standard SMOTE, and random oversampling with synthetic-profile auditing and repeated validation, while explicitly distinguishing operational score-category reproduction from independently validated psychological prediction. Originality – This study combines a controlled comparison of baseline Random Forest, standard SMOTE, and random oversampling with synthetic-profile auditing and repeated validation, while explicitly distinguishing operational score-category reproduction from independently validated psychological prediction.]]>
<![CDATA[KLASIFIKASI KERENTANAN PHISHING DI KALANGAN MAHASISWA MENGGUNAKAN DECISION TREE ]]>
<![CDATA[Mhd. Murini Ramadhani]]>;
<![CDATA[Muhammad Ikhsan]]>
<![CDATA[ Journal of Digital Technology and Computer Science Vol. 3 No. 3 (2026): August 2026 ]]>
Publisher : <![CDATA[Academic Bright Collaboration ]]>
Show Abstract
|
Download Original
|
Original Source
|
Check in Google Scholar
|
DOI: 10.66053/dtcs.v3i3.1165
<![CDATA[Purpose – This study aims to classify phishing vulnerability among college students and identify the most influential factors using the C4.5 Decision Tree algorithm. Methods – A quantitative survey was conducted among 171 active Computer Science students at the State Islamic University of North Sumatra, Medan, Indonesia, from the 2023–2025 cohorts. Behavior, Knowledge, and cybersecurity training experience were used as predictor variables, while the vulnerability labels Vulnerable, Alert, and Aware were derived from responses to four hypothetical phishing scenarios. The model was developed using RapidMiner and evaluated through 10-fold cross-validation, a confusion matrix, and multiclass Receiver Operating Characteristic Area Under the Curve (ROC-AUC) using the One-vs-Rest approach. Findings – The dataset consisted of 131 Aware students (76.61%), 28 Alert students (16.37%), and 12 Vulnerable students (7.02%). The primary C4.5 model achieved 85.96% aggregate accuracy, but performance was uneven across classes: the macro-average F1-score was 69.84%, the Alert-class F1-score was 46.51% with 35.71% recall, and the Aware-class F1-score was 93.77%. The macro-average AUC was 0.8263. Phishing knowledge was the dominant predictor and formed the root node of the decision tree. Research implications – The model performed strongly for the Aware class but inconsistently across classes. The findings may inform exploratory cybersecurity education planning, but interpretation should remain cautious because of severe class imbalance, sensitivity to the operational vulnerability cutoffs, and recruitment from a single study program. Originality – The contribution is primarily contextual and application-oriented: phishing vulnerability is operationalized from scenario responses and examined together with behavioral, knowledge, and training factors in an interpretable multiclass model for an educational cybersecurity context, rather than as a methodological innovation in the C4.5 algorithm.]]>
<![CDATA[Comparing Decision Tree and SVM with Hyperparameter Tuning for Classifying Perceived Social Media Impact on Academic Performance ]]>
<![CDATA[Nur Sitorus]]>;
<![CDATA[Yusuf Ramadhan Nasution]]>
<![CDATA[ Journal of Digital Technology and Computer Science Vol. 3 No. 3 (2026): August 2026 ]]>
Publisher : <![CDATA[Academic Bright Collaboration ]]>
Show Abstract
|
Download Original
|
Original Source
|
Check in Google Scholar
|
DOI: 10.66053/dtcs.v3i3.1173
<![CDATA[Purpose – This study compares the performance of Decision Tree and Support Vector Machine (SVM) in classifying students’ perceived academic impact of social media use and examines the contribution of Grid Search-based hyperparameter tuning to model performance. Methods – A quantitative experimental approach was applied to a subset of records obtained from a publicly available Kaggle dataset. The records were retained in their original sequence. Data preprocessing included data quality inspection, binary target encoding, One-Hot Encoding for categorical variables, and feature scaling for numerical variables used in the SVM model. The dataset was sequentially divided into training, validation, and testing subsets without shuffling or stratification. Grid Search was performed using a predefined validation split, while model performance was assessed using accuracy, weighted precision, weighted recall, and weighted F1-score. Findings – Hyperparameter tuning improved the overall performance of the Decision Tree model compared with its default configuration. In contrast, the default and optimized SVM models produced identical results because Grid Search selected the same parameter configuration as the default settings. Among the evaluated models, SVM achieved the highest observed testing performance. Research implications – The findings indicate that the effectiveness of hyperparameter tuning may vary across machine learning algorithms. Grid Search improved the Decision Tree configuration while confirming the suitability of the default SVM settings within the evaluated search space. Since the target variable reflects students’ perceived academic impact rather than objectively measured academic achievement, the results should be interpreted within this specific classification context and the limitations of a single public dataset. Originality – This study provides a controlled comparison of default and optimized Decision Tree and SVM models under identical experimental conditions, highlighting algorithm-specific responses to Grid Search-based hyperparameter tuning.]]>
<![CDATA[Clustering of Student Personality Types Based on Extracurricular Activities using The K-Means Algorithm Approach ]]>
<![CDATA[Rifdah Syahputri]]>;
<![CDATA[Ilka Zufria]]>
<![CDATA[ Journal of Digital Technology and Computer Science Vol. 3 No. 3 (2026): August 2026 ]]>
Publisher : <![CDATA[Academic Bright Collaboration ]]>
Show Abstract
|
Download Original
|
Original Source
|
Check in Google Scholar
|
DOI: 10.66053/dtcs.v3i3.1267
<![CDATA[Purpose – This study aims to identify patterns of student personality tendencies at Universitas Islam Negeri Sumatera Utara (UINSU) using a data-driven clustering approach based on extracurricular participation characteristics and personality-related questionnaire responses. The study was motivated by the need to systematically analyze the relationships and patterns among students' organizational activity, activity frequency, organizational roles, social interaction, and personality tendency characteristics. Methods – This study employed a quantitative approach using data mining. Data were collected through a questionnaire distributed to UINSU students, resulting in 604 respondents. The data were processed through cleaning, missing-value checking, transformation, and Min-Max normalization. The K-Means Clustering algorithm was applied to group students according to their extracurricular activity patterns. The clustering quality was evaluated using the Silhouette Score, while PCA was used to visualize the clustering results. Findings – The clustering process identified three personality tendency groups: Extrovert, Introvert, and Ambivert. The resulting clustering obtained a Silhouette Score of 0.3823, indicating a moderate level of cohesion and separation among the clusters, although some observations remained relatively close to other clusters. Research implications – The findings provide an overview of student personality tendencies based on extracurricular activity patterns. However, the results represent personality tendencies derived from clustering characteristics and should not be interpreted as psychological diagnoses. Originality – This study applies K-Means Clustering to extracurricular activity data as a basis for identifying student personality tendencies into Introvert, Extrovert, and Ambivert groups, providing a data-driven perspective for understanding student characteristics.]]>
<![CDATA[Integrated Web Application Firewall with AbuseIPDB, Iptables, and Telegram Notifications for Real-Time Web Security Monitoring ]]>
<![CDATA[Robby Maulana]]>;
<![CDATA[Imam Sutanto]]>;
<![CDATA[Nizirwan Anwar]]>;
<![CDATA[Ari Pambudi]]>
<![CDATA[ Journal of Digital Technology and Computer Science Vol. 3 No. 3 (2026): August 2026 ]]>
Publisher : <![CDATA[Academic Bright Collaboration ]]>
Show Abstract
|
Download Original
|
Original Source
|
Check in Google Scholar
|
DOI: 10.66053/dtcs.v3i3.1096
<![CDATA[Purpose – This study designs and evaluates an integrated web-security workflow combining ModSecurity, AbuseIPDB, Iptables, and Telegram to connect attack detection, IP-reputation assessment, conditional firewall-rule management, and real-time administrator decisions. Methods – A Research and Development approach was applied in a controlled Ubuntu 24.04 laboratory using Apache 2.4.58, OWASP Core Rule Set, Python 3.12.3, a PHP test application, AbuseIPDB API, Telegram API, and Iptables. Eighteen payloads, comprising six cross-site scripting, six SQL injection, and six local file inclusion cases, were each executed once under two reputation conditions defined by an Abuse Confidence Score threshold of 75. Findings – ModSecurity rejected and logged 18 of 18 selected malicious requests with HTTP 403 responses. All 9 of 9 high-risk cases produced the expected Iptables rule insertion, while all 18 detected events generated Telegram notifications with appropriate Blocklist or Whitelist controls. These results confirm completion of the configured workflow but not subsequent packet rejection from verified network-layer sources. Research implications – Generalizability is limited by the small attack-only dataset, single execution per payload, one laboratory environment, absence of legitimate-traffic controls and advanced evasion tests, reliance on X-Forwarded-For values, and lack of follow-up connections. Consequently, specificity, false-positive rate, production reliability, and actual network containment remain unmeasured. Originality – This study extends prior ModSecurity and AbuseIPDB integration by adding reputation-based Iptables rule insertion and interactive Telegram rule management within one reproducible, human-in-the-loop response chain.]]>
<![CDATA[Reka: Asisten AI Marketing Generatif Berbasis WhatsApp untuk UMKM Indonesia Menggunakan Pendekatan Design Science Research ]]>
<![CDATA[Satrio Tegar Sadewo]]>;
<![CDATA[Reni Eka Septiani]]>;
<![CDATA[Aditya Kumala Dewi]]>
<![CDATA[ Journal of Digital Technology and Computer Science Vol. 3 No. 3 (2026): August 2026 ]]>
Publisher : <![CDATA[Academic Bright Collaboration ]]>
Show Abstract
|
Download Original
|
Original Source
|
Check in Google Scholar
|
DOI: 10.66053/dtcs.v3i3.1131
<![CDATA[Purpose – Micro, Small, and Medium Enterprises (MSMEs) in Indonesia face difficulty producing digital marketing content consistently due to limited human resources and digital literacy, while existing generative AI applications still demand high technical skills and switching between platforms. This study aims to design, build, and evaluate "Reka," a WhatsApp-based AI marketing assistant that integrates generative AI for text and images into a single end-to-end conversation flow. Method – The study uses a Design Science Research approach. The system was built on the n8n architecture and Google Gemini with a rule-based NLU pipeline to handle natural-language input across seven functional modules. The evaluation involved 27 respondents who ran eleven usage scenarios through Black Box Testing, the System Usability Scale (SUS), and the Perceived Content Quality Evaluation (PCQ). Findings – The evaluation showed a Black Box Testing weighted success rate of 85.4%, an SUS score of 68.80 (Grade C/Marginal Acceptable), and a PCQ score of 3.81 out of 5.00. Nine semantic bug cases were also found, categorized into three patterns. Research Implications – Reka demonstrates initial feasibility as a WhatsApp-based AI marketing assistant prototype in a controlled testing environment, with noted room for improvement in visual quality and the content-revision feature. Further validation with actual MSME operators is still needed, given that the respondents in this study were students role-playing as owners of fictitious MSMEs. Originality – This study integrates generative AI for text and images into a single end-to-end WhatsApp conversation flow to produce MSME marketing content, encompassing seven functional modules within one system.]]>
<![CDATA[SENTIMEN ANALISIS TERHADAP UJARAN KEBENCIAN ISU POLITIK DI INDONESIA DENGAN HYBRID MODEL MACHINE LEARNING ]]>
<![CDATA[Chandra Haris Mulyana]]>;
<![CDATA[Sawali Wahyu]]>
<![CDATA[ Journal of Digital Technology and Computer Science Vol. 3 No. 3 (2026): August 2026 ]]>
Publisher : <![CDATA[Academic Bright Collaboration ]]>
Show Abstract
|
Download Original
|
Original Source
|
Check in Google Scholar
|
DOI: 10.66053/dtcs.v3i3.1137
<![CDATA[Purpose - This study evaluates an auditable hybrid machine-learning workflow for sentiment classification and operational hate-speech detection in Indonesian political YouTube comments, distinguishing agreement with lexicon-derived labels from agreement with human judgments. Methods - From 25,651 comments, 25,484 were retained. The sentiment target was reconstructed from a 10,272-row polarity lexicon, and the operational target was audited against a 545-term lexicon. Unigram-bigram TF-IDF and lexical features were tested using Multinomial Naive Bayes, linear SVM, Random Forest, and hard voting. Evaluation used an 80:20 holdout, five-fold validation, ablation, keyword masking, McNemar tests, a blinded 400-comment dual-annotation reference set with adjudication, and out-of-fold regional prediction. Findings - Linear SVM achieved weighted F1 scores of 83.29% for sentiment and 99.03% for the operational hate-speech target. On the human reference set, SVM without hate-speech features achieved 66.50% accuracy and 18.29% hate-class F1; the operational rule achieved 18.93% hate-class F1. Annotator agreement was 82.25% with Cohen's kappa of 0.1834. Neutral sentiment remained the weakest class. Regional estimates differed by at most 1.93 percentage points between SVM and voting. Research implications - Near-ceiling operational-target performance does not establish general moderation accuracy. The workflow supports exploratory screening and human-review prioritization. Regional outputs describe source-video topic groups, not provincial populations. Originality - The study integrates target reconstruction, leakage-aware testing, human-reference evaluation, model comparison, and out-of-fold regional visualization in one reproducible workflow.]]>
