Journal Of Data Insights

Sentiment Analysis of FlyGaruda Review Using Support Vector Machine and Naive Bayes Algorithm

Gibran Masta Pangestu Baskoro, Galet Guntoro Setiaji, Ahmad Rifa’i — Tue, 30 Jun 2026 00:00:00 +0000

FlyGaruda is an official digital application owned by Garuda Indonesia that provides ticket booking and online check-in services for users. This study analyzed the sentiment of reviews on the Google Play Store by comparing the performance of Support Vector Machine and Multinomial Naive Bayes. The methods used include scraping, text preprocessing, extraction of the Term Frequency-Inverse Document Frequency (TF-IDF) feature, and evaluation using the Confusion Matrix. The dataset used totaled 4,790 reviews with positive, negative, and neutral categories. The results showed that both models obtained an accuracy of 82.25%. However, the Support Vector Machine produces a weighted precision of 77.66% and an F1-Score of 78.91%, better at handling data imbalances. Meanwhile, Multinomial Naive Bayes excels in computing efficiency with a training time of 0.08 seconds compared to 90.60 seconds on the Support Vector Machine. In conclusion, although it is slower, the Support Vector Machine provides more consistent and accurate classification performance. This research contributes to the development of a machine learning-based opinion analysis system to improve the quality of aviation digital services in a sustainable manner. These findings can serve as a reference in the selection of the best algorithms between accuracy and computational speed in large text data and support data-driven decision-making in the modern air transportation industry in the current era of global sustainable digital transformation

K-Nearest Neighbor Algorithm in Classification of Stunting Detection Dataset

Tue, 30 Jun 2026 00:00:00 +0000

Stunting is a nutritional problem that can affect children's physical growth and cognitive development and has a long-term impact on the quality of future generations. Early detection of stunting is crucial to enable timely and effective interventions. As technology advances, machine learning algorithms such as K-Nearest Neighbors (KNN) offer potential solutions to improve the accuracy of stunting risk classification. This study aims to design a classification model based on the K-Nearest Neighbors (KNN) algorithm in the early detection of stunting risk in toddlers. This research uses the 2024 stunting dataset obtained from Kaggle. The data is analyzed through the stages of cleaning, transformation, and division into training and testing data. The KNN model was tested with various K values to determine the optimal value. The results showed that the KNN model with a value of K=8 resulted in an accuracy of 93.80%, F1-Score of 93.65%, precision of 93.63%, and recall of 93.79%. This shows that KNN is reliable in classifying the nutritional status of toddlers and can be applied in stunting prevention efforts using more accurate data. This research contributes to developing machine learning-based classification systems that can support decision-making in public health programs, especially in reducing stunting rates.

Comparison of Decision Tree and Random Forest Performance for Sentiment Analysis of Public Service App Reviews

Aditya Rizky Purnama, Galet Guntoro Setiaji, Ahmad Rifa'i — Tue, 30 Jun 2026 00:00:00 +0000

The Info BMKG application is a government-developed digital public service platform designed to provide real-time weather, seismic, and climate information to Indonesian citizens. The substantial volume of user reviews accumulated on the Google Play Store holds significant potential as a service evaluation resource; however, the limitations of manual review processes necessitate an efficient computational approach. This study proposes a machine learning-based sentiment analysis framework to classify user reviews of the Info BMKG application, while systematically comparing the performance of two algorithms Decision Tree and Random Forest using a dataset of 10,000 reviews collected via web scraping. The data underwent text preprocessing, rating-based sentiment labeling, and TF-IDF feature extraction, followed by evaluation using accuracy, precision, recall, F1-score, cross-validation, and computational time metrics. Experimental results demonstrate that Random Forest achieved 81% accuracy with a 77% F1-score, outperforming Decision Tree which attained 78% accuracy and 75% F1-score. In terms of efficiency, Decision Tree exhibited faster testing time (0.114 seconds) compared to Random Forest (0.201 seconds), while Random Forest proved more efficient in training time (7.347 seconds versus 12.421 seconds). These findings confirm that Random Forest represents the more optimal algorithm for sentiment classification tasks involving public service application user reviews.

ADASYN-Based Multiclass Support Vector Machine for Village Development Index Classification in North Maluku Province

Tiani Wahyu Utami, Lea Angelina, Saeful Amri — Tue, 30 Jun 2026 00:00:00 +0000

Class imbalance is a significant constraint that can diminish the performance of classification models. This study implements the integration of Adaptive Synthetic Sampling (ADASYN) and Multiclass Support Vector Machine (SVM) to classify the 2024 Village Development Index (IDM) in North Maluku Province. The dataset comprises 684 villages, utilizing the Social Resilience Index (IKS), Economic Resilience Index (IKE), and Environmental Resilience Index (IKL) as predictor variables. The data was partitioned using a ratio of 80% for training and 20% for testing. An extreme imbalance was identified in the "independent village" category (0.88%); therefore, ADASYN was applied to the training data to generate 862 synthetic samples to balance the class distribution. The optimal model yielded by the process was a linear kernel SVM with a Cost value of 100, yielding an accuracy of 98.54%, precision of 98.26%, recall of 99.4%, and an F1-score of 98.83%. Of the total 137 villages evaluated, only two villages were misclassified: Salimuli Village and Dowongimaiti Village. These findings demonstrate the effectiveness of the ADASYN-SVM combination in producing accurate classifications to support village development policies in island regions.

Development of a Decision Support System for Regional Competitiveness Policy Recommendations Based on Explainable Artificial Intelligence (XAI)

Sintha Istikomah, Dwi Purnomo Putro; Sholihul Ibad; Aditya Hermawan — Tue, 30 Jun 2026 00:00:00 +0000

Enhancing regional competitiveness is a critical factor in driving economic growth, investment, and community welfare. However, the utilization of Regional Competitiveness Index (Indeks Daya Saing Daerah/IDSD) data in Indonesia has largely been limited to ranking purposes, thus failing to provide specific, data-driven policy recommendations. This study aims to develop a Decision Support System (DSS) for regional competitiveness policy recommendations by combining machine learning and Explainable Artificial Intelligence (XAI) within a Design Science Research (DSR) framework. The dataset originates from provincial IDSD data spanning 2022–2025, encompassing 12 assessment pillars as predictor variables. Three regression algorithms were examined: Linear Regression, Random Forest, and XGBoost. A Variance Inflation Factor (VIF) analysis was conducted to verify the absence of severe multicollinearity among the predictor variables. Based on performance evaluation, XGBoost was selected as the final model due to its superior predictive performance and stability, yielding an R² of 0.8712 on the 2025 test data and a mean 5-fold cross-validation R² of 0.7723. To enhance model transparency, SHapley Additive exPlanations (SHAP) was employed. Interpretation results revealed that Innovation Capability (Pillar 12), Adoption of Information and Communication Technology (Pillar 3), and Market Size (Pillar 10) are the most influential factors affecting regional competitiveness scores. Building on these findings, the developed system delivers context-specific, priority policy recommendations through an interactive dashboard. This study demonstrates that the integration of XGBoost and XAI constitutes a more objective, transparent, and adaptive data-driven decision-making solution for supporting regional competitiveness improvement in Indonesia.

Hate Speech Detection on X Using K-Nearest Neighbor with TF–IDF and Cosine Similarity

Tue, 30 Jun 2026 00:00:00 +0000

The rapid growth of social media has increased online interactions but has also accelerated the spread of hate speech content that may negatively impact individuals and communities. X (formerly Twitter), as one of the largest social networking platforms, enables users to share opinions publicly, making automatic hate speech detection increasingly important. This research proposes a hate speech classification approach using the K-Nearest Neighbor (KNN) algorithm combined with Term Frequency–Inverse Document Frequency (TF–IDF) weighting and Cosine Similarity. The dataset consists of 900 social media posts collected through the platform API and manually labeled into hate speech and non-hate speech categories, consisting of 675 training data and 225 testing data. Prior to classification, text preprocessing techniques including tokenization, stopword removal, and stemming were applied to improve text quality. Model evaluation was conducted using 10-fold cross validation to assess classification performance. Experimental results showed that the KNN algorithm with Cosine Similarity distance measurement and K=3 parameter achieved an accuracy of 78.22% in hate speech detection tasks. The findings indicate that KNN combined with TF–IDF and Cosine Similarity provides a reliable approach for social media text classification and can support automated hate speech detection systems.

Machine Learning Based Water Level Forecasting for Tidal Flood Mitigation in Indonesian Coastal Regions

Sugiarto Sugiarto, Riduan Riduan, Linus Yoseph Wawan Rukmono — Tue, 30 Jun 2026 00:00:00 +0000

Changes in coastal water levels significantly impact maritime logistics and tidal flood risks, which are categorised as non-military threats to national security and socio-geographical stability. This study aims to develop and compare water level prediction models using Random Forest, Long Short-Term Memory (LSTM), and their ensemble as a technology-based disaster mitigation strategy in Indonesian coastal areas. The dataset was obtained from the Tanjung Priok Maritime AWS in August 2024 and comprises meteorological features (rainfall, air temperature, air pressure, wind speed) and oceanographic data (sea surface temperature), with water level as the target variable. Data preprocessing involved time-based linear interpolation and feature standardisation. Model evaluation was conducted using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and residual analysis. The results indicate that Random Forest provides stable performance (RMSE 0.2957, MAE 0.2130), while LSTM is more responsive to daily temporal fluctuations despite slightly higher errors (RMSE 0.3156, MAE 0.2339). A simple ensemble of both models achieved the most optimal and robust performance (RMSE 0.2954, MAE 0.2132) with a well-distributed residual. This study concludes that implementing the ensemble model enhances the reliability of tidal flood early warning systems, serving as a crucial pillar of non-military defence strategies to safeguard national logistical security in Indonesia's coastal zones.

A Hybrid Decision Support System for Rice Plant Disease Diagnosis and Treatment Recommendation Using Dempster-Shafer, AHP-TOPSIS, and Fuzzy SAW

Tue, 30 Jun 2026 00:00:00 +0000

Rice diseases — blast (Magnaporthe oryzae), bacterial leaf blight (Xanthomonas oryzae pv. oryzae), and sheath blight (Rhizoctonia solani)—cause annual global yield losses of 10–100%, resulting in billions of U.S. dollars in economic damage. Smallholder farmers in remote regions often lack access to agronomy experts and face difficulties using image-based diagnostic systems on low-capacity devices. This study proposes and evaluates a hybrid three-module Decision Support System (DSS) framework based on non-image tabular data to address these challenges. The framework integrates: (1) Dempster–Shafer Theory for probabilistic disease diagnosis using 48 structured clinical symptom parameters from ESforRPD2; (2) a hybrid AHP–TOPSIS module with CRITIC-based objective weight verification for multicriteria treatment ranking; and (3) an adaptive Fuzzy SAW module employing dynamic weights based on crop growth stages derived from Paddy Doctor Metadata. Experimental results show that the Dempster–Shafer module achieved 88.9% accuracy, a macro F1-score of 0.877, and a macro AUC-ROC of 0.939, outperforming Certainty Factor (82.4%), Random Forest (85.7%), and XGBoost (86.1%). The AHP model produced a valid Consistency Ratio (CR = 0.030), while CRITIC analysis revealed substantial differences between expert-assigned and data-driven weights. The adaptive Fuzzy SAW module achieved 100% agreement with agronomy expert recommendations (Spearman’s rho = 0.941), surpassing static SAW (25%, rho = 0.487) and standalone TOPSIS (0%, rho = 0.412). The framework operates without image input and provides recommendations in under two seconds, making it suitable for low-capacity devices and remote agricultural environment

The Development of Artificial Intelligence in Defense Command and Control (C2) Systems A Literature Review

Ahmad Fajrin Kusuma Wijaya, Riduan Riduan — Tue, 30 Jun 2026 00:00:00 +0000

This study analyzes developments in artificial intelligence (AI) for defense command and control (C2) systems through an in-depth synthesis of 25 Scopus-indexed international journals (10 Q1, 8 Q2, and 7 Q3) published between 2021 and 2023. The study identified six major AI technology categories that dominate defense C2 research: Decision Support Systems (24%), Explainable AI & Trust (24%), Situational Awareness (16%), Machine Learning & Deep Learning (12%), Multi-Agent Systems (12%), and Security & Risk Management (12%). The research gaps analysis revealed critical challenges in legacy system integration, standardization of explainability metrics, AI adaptation to dynamic adversary tactics, management of operator cognitive load, implementation of an ethical framework, and resilience against adversarial attacks. This research found that while technologies such as Deep Reinforcement Learning and Multi-Agent Systems have reached Technology Readiness Level (TRL) 6-8 (approaching the operational stage), Human-Autonomy Teaming implementations are still at TRL 3-5, indicating significant further research needs. The analysis also shows a sharp increase in publication trends, from 1 in 2021 to 13 in 2023 (an ~1300% increase), reflecting the rapidly increasing global research intensity. This study recommends developing hybrid frameworks for federated learning, military-domain-specific explainable AI techniques, multi-agent reinforcement learning algorithms with transfer learning, and AI accountability mechanisms integrated with international humanitarian law as future research priorities. The findings and recommendations are expected to support the academic community, military practitioners, and policymakers in accelerating the responsible and effective adoption of defense C2 AI.

A Classification of Debunking in Indonesian Fact-Checking Platforms Using NLP and Machine Learning : A Mixed-Methods Approach with Corpus Analysis and IndoBERT

Bayu Hartono, Riduan Riduan, Rudy Agus Gemilang Gultom, Hondor Saragih — Tue, 30 Jun 2026 00:00:00 +0000

The rapid spread of disinformation through digital platforms constitutes a serious threat to social cohesion and public health. Debunking—the systematic refutation of false information using verified evidence—has emerged as a key countermeasure, yet manual identification and classification of debunking strategies is labor-intensive and difficult to scale. This study addresses this gap through a mixed-methods design integrating qualitative corpus analysis with automated machine learning (ML) classification. A corpus of 120 debunking articles published by three leading Indonesian fact-checking institutions (Kominfo AIS, Mafindo, and Cek Fakta Kompas, 2022–2024) was first manually annotated by two trained coders (Cohen's κ = 0.82) to identify four dominant debunking strategies: (1) contextual correction with emotional narrative framing; (2) source authority endorsement; (3) visual verification and reverse image search; and (4) myth-versus-fact inoculation format. This annotated corpus was subsequently used as a training dataset to develop and benchmark five NLP-based text classification models: TF-IDF + Support Vector Machine (SVM), TF-IDF + Random Forest, IndoBERT fine-tuned, IndoBERT with data augmentation (IndoBERT-Aug), and XGBoost with linguistic features. The IndoBERT-Aug model achieved the highest overall performance (macro-averaged F1 = 0.847, Precision = 0.851, Recall = 0.843), substantially outperforming the SVM baseline (F1 = 0.612). Logistic regression analysis further identified three significant moderators of debunking effectiveness: correction timeliness within 6 hours (OR=2.80, p<0.01), content readability (OR=0.68, p<0.01), and multi-platform distribution (OR=1.84, p<0.05), with the full model explaining 41% of variance (Nagelkerke R²=0.41). These contributions are formalized into the Indonesian Debunking Effectiveness Model (IDEM), a framework integrating automated strategy detection with evidence-based deployment guidelines for scalable counter-disinformation operations.

The AI-Enabled Pharmacovigilance for Defence Health Surveillance: Automatic Detection of Adverse Drug Events from Patient Reviews Using BioClinical ModernBERT

Tue, 30 Jun 2026 00:00:00 +0000

Pharmacovigilance is a critical component of post-marketing drug safety, yet conventional adverse drug event (ADE) reporting systems remain constrained by substantial underreporting. In defence health systems, the timely detection of medication-related safety signals is not only a clinical concern but also a matter of force health protection, medical readiness, and operational resilience. Patient-generated health narratives from online forums, drug review platforms, and social media provide a complementary source of pharmacovigilance intelligence, but their informal, unstructured, and context-dependent nature creates significant challenges for automated analysis. This study evaluates BioClinical ModernBERT, a biomedical–clinical long-context encoder based on the ModernBERT architecture, for automatic ADE detection from patient reviews. Its performance is compared with three representative BERT-based transformer baselines: BERT-base, BioBERT, and ClinicalBERT. Experiments were conducted using the CSIRO Adverse Drug Event Corpus (CADEC), a benchmark corpus of patient-reported medication experiences from online health forums. The corpus was pre-processed through sentence segmentation, text cleaning, medical entity normalization, and sentence-level label alignment for binary ADE/non-ADE classification. All models were fine-tuned using a 70:15:15 training, validation, and test split and evaluated using accuracy, precision, recall, and F1-score. The results show that BioClinical ModernBERT achieved the highest overall performance, with an F1-score of 0.891, outperforming ClinicalBERT (0.847), BioBERT (0.832), and BERT-base (0.798). Confusion matrix analysis further indicates that BioClinical ModernBERT reduced false negative errors, particularly in long, multi-clause, and clinically implicit patient narratives. These findings suggest that combining biomedical–clinical domain adaptation with long-context representation provides a meaningful advantage for detecting ADE signals in complex patient-generated text. From a defence health perspective, such capability may support the development of AI-enabled pharmacovigilance surveillance systems that enhance medication safety, health intelligence, and readiness-oriented risk monitoring across civilian–military health ecosystems.

MSME Digital Transformation Readiness Prediction for Data-Driven Decision Support

Andi Riansyah, Maya Indriastuti, Maulana Ahmad Widiarta — Tue, 30 Jun 2026 00:00:00 +0000

Digital transformation readiness among micro, small, and medium enterprises (MSMEs) varies across business sectors and organizational capabilities. This study develops a supervised classification model to predict MSME digital transformation readiness for data-driven decision support. Target labels were constructed from previous fuzzy clustering results and organized into three readiness levels: low, moderate, and high. The predictive model was built using Random Forest with business type, workforce transformation, dynamic capability, and SME performance as key predictors. Data preprocessing included categorical encoding, train-test separation, and class imbalance handling applied only to the training data to avoid data leakage. Model evaluation on the hold-out set produced 91.40% accuracy, 91.36% macro precision, 92.16% macro recall, and 91.72% macro F1-score. The confusion matrix showed that 85 of 93 test observations were correctly classified, with most errors occurring between adjacent readiness levels. Feature importance analysis indicated that dynamic capability was the most influential predictor, followed by workforce transformation and SME performance. The findings demonstrate that Random Forest can transform clustering-based insights into a practical predictive model for prioritizing MSME assistance, training, and digital development programs.

CoVaR Modelling using QRNN Based on Quantile Regression And Quantile Autoregressive Models with Stochastic Search Variable Selection for LQ45

Zulfa Wahyu Mardika, Dedy Dwi Prastyo, T. Dwi Ary Widhianingsih — Tue, 30 Jun 2026 00:00:00 +0000

Rapid fluctuations in stock prices, particularly during periods of market turmoil, can increase the risk of extreme losses and trigger risk contagion across firms. In risk management practice, Value-at-Risk (VaR) is widely used to measure potential losses at the individual asset or portfolio level. However, VaR is not sufficient to explain how the risk of a given firm changes when another firm or the market is under distress. To address this limitation, Conditional Value-at-Risk (CoVaR) is employed to measure the risk of a firm conditional on extreme conditions affecting another firm or the market, making it more relevant for representing systemic risk contributions and spillover effects among highly liquid stocks such as those included in the LQ45 index. Accordingly, this study focuses on optimizing the input variables of the CoVaR model for the returns of firms included in the LQ45 index by integrating Quantile Regression Neural Network (QRNN) as a nonlinear quantile model and Stochastic Search Variable Selection (SSVS) as a Bayesian variable selection mechanism based on posterior inclusion probability. Within this framework, VaR is first estimated dynamically using Quantile Autoregressive (QAR) and subsequently used as a reference for the distress condition in the CoVaR model. CoVaR is then modelled using QRNN, while the candidate input variables are optimized using SSVS. QRNN is chosen because it is capable of modelling extreme quantiles when the relationship between returns and risk factors is not necessarily linear and tends to vary with market conditions, whereas SSVS is employed to obtain more parsimonious inputs, reduce multicollinearity, mitigate the risk of overfitting, and improve the interpretability of dominant factors.

A Hybrid AHP-TOPSIS Decision Support System with Temporal Ridge Regression for Dynamic Prediction of Regional Food Vulnerability Index

Surya Supratman, Riduan Riduan, Mokhamad Solikin, Maulana Muhammad Jogo Samodro — Tue, 30 Jun 2026 00:00:00 +0000

Regional food vulnerability in Indonesia is a dynamic and multidimensional challenge that requires timely and accurate monitoring. However, the annual Food Security and Vulnerability Atlas (FSVA) remains limited in its ability to capture rapid intra-annual changes in food security conditions, reducing its effectiveness as an early-warning instrument. This limitation became evident during the 2023 El Niño event, which caused significant production shocks that were not reflected in official vulnerability assessments until the following year. This study proposes a Hybrid Multi-Criteria Decision-Making (HMCDM) framework integrating the Analytic Hierarchy Process (AHP), Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), and Ridge Regression to generate a dynamic Regional Food Vulnerability Index (RFVI). The framework was evaluated using a 36-month panel dataset covering 30 sub-districts and nine food security indicators. Expert-derived criteria weights were validated through AHP consistency testing (CR = 0.056), while monthly TOPSIS scores were transformed into supervised learning targets using a novel TOPSIS-as-ML-target architecture. Temporal prediction was performed using Ridge Regression with lag-based feature engineering and expanding-window cross-validation. The proposed model achieved strong predictive performance ((R^2 = 0.870), MAE = 0.043, RMSE = 0.061), outperforming standalone Ridge Regression, ARIMA, and Naïve Forecast baselines. Vulnerability classification accuracy reached 97.3%, while Spearman correlation analysis ((\rho = 0.831), (p < 0.01)) confirmed substantial agreement between expert-defined priorities and data-driven feature importance. The results demonstrate that integrating multicriteria evaluation with temporal machine learning can significantly improve food vulnerability forecasting. The proposed framework provides a robust foundation for data-driven early-warning systems and proactive food security policy planning.

A Hybrid LSTM with GARCH-MIDAS-X for Modelling IDX Composite Volatility

Silviya Indriyani, Irhamah, Tintrim Dwi Ary Widhianingsih — Tue, 30 Jun 2026 00:00:00 +0000

Stock market volatility forecasting plays a crucial role in supporting investment decision-making and risk management under uncertain market conditions. This study proposes a hybrid LSTM with GARCH-to modelling IDX Composite volatility. The GARCH-MIDAS-X model is first employed to decompose stock return volatility into short-run and long-run components while incorporating multiple low-frequency exogenous variables, including market news sentiment, crude oil prices, and exchange rates. The residual generated by the GARCH-MIDAS-X model is subsequently used as input for the LSTM network to capture complex nonlinear patterns and temporal dependencies that may not be fully explained by the econometric model. Model performance is evaluated through both in-sample and out-of-sample forecasting using several accuracy measures, including Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE). The empirical results indicate that the hybrid model produces forecasting performance comparable to that of the GARCH-MIDAS-X model, with only marginal differences in prediction accuracy. These findings suggest that the GARCH-MIDAS-X model is capable of capturing most of the relevant volatility dynamics, while the addition of the LSTM component provides limited incremental forecasting benefits for the observed period. Therefore, the hybrid approach may serve as an alternative forecasting framework, although its superiority over the standalone econometric model is not evident in this study.

Systematic Review of Multimodal Emotion Recognition in the Wild: Integrating Facial Expressions, Speech, and Physiological Signals for Enhanced Context-Aware Applications

Muhammad Munsarif, Norshuhani Zamin, Richmond Ampah-Mensah — Tue, 30 Jun 2026 00:00:00 +0000

Facial Emotion Recognition (FER) has become a critical component of affective computing, human–computer interaction, intelligent healthcare, adaptive education, and assistive technologies. This systematic literature review synthesizes recent developments in multimodal emotion recognition in the wild by examining how facial expressions, speech, physiological signals, deep learning architectures, and deployment technologies shape robust context-aware FER systems. Following the PRISMA protocol, literature was identified from Scopus using FER-related deep learning keywords, resulting in 202 initial records and 61 eligible studies for thematic synthesis, trend analysis, methodological classification, and qualitative interpretation. The findings show that FER research has shifted strongly from handcrafted features toward CNN-based deep learning, transfer learning, hybrid architectures, attention mechanisms, and bio-inspired optimization. Human–computer interaction emerged as the dominant research context, while healthcare, autism spectrum disorder screening, education, assistive technology, mining safety, and smart services represent increasingly important application domains. Transfer learning dominated robustness strategies, while multimodal fusion using facial images, speech, EEG, wearable sensors, and audio-visual signals gained stronger academic attention because it improves contextual understanding and reduces the limitations of unimodal FER. The synthesis also reveals persistent challenges, including poor generalization in uncontrolled environments, dataset imbalance, cultural variation, micro-expression recognition, computational complexity, real-time deployment, and limited explainability. The review contributes a multidimensional conceptual perspective that integrates multimodal sensing, deep learning optimization, edge/IoT deployment, and ethical-aware application design. Future research should prioritize lightweight multimodal FER, cross-cultural datasets, explainable AI, privacy-preserving learning, real-world clinical validation, and adaptive systems capable of operating reliably under noisy, dynamic, and socially sensitive conditions.

Hyperparameter Optimization of Random Forest using Grey Wolf Optimization for Heart Disease Classification

Ratih Khotimahtus Sa'diyah, Muhammad Sam’an, Safuan, Mustafa Mat Deris — Tue, 30 Jun 2026 00:00:00 +0000

Cardiovascular disease remains one of the leading causes of death worldwide, making predictive models important to support early heart disease detection. Random Forest is widely used for heart disease classification, but its performance can be affected by hyperparameter selection. This study focuses on applying Grey Wolf Optimization (GWO) to selected Random Forest hyperparameters and evaluating the optimized model through a direct comparison with a baseline Random Forest model on the same testing dataset, supported by statistical verification. The dataset used is the Cleveland Heart Disease Dataset, consisting of 303 patient records, 13 predictor attributes, and one target attribute. The research stages include data preparation, preprocessing, stratified data splitting with an 80:20 ratio, hyperparameter optimization using GWO, and model evaluation. The GWO process uses the average F1-score from 5-fold cross-validation on the training set as the fitness value. Model performance is evaluated using accuracy, precision, recall, F1-score, AUC-ROC, confusion matrix analysis, and the exact McNemar test. The results show that the GWO-RF model obtains higher descriptive evaluation values than the baseline RF model, with accuracy increasing from 88.52% to 93.44%, precision from 81.82% to 90.00%, F1-score from 88.52% to 93.10%, and AUC-ROC from 95.13% to 96.86%, while recall remains at 96.43%. However, the exact McNemar test produces a p-value of 0.25, indicating that the difference is not statistically significant. Therefore, the improvement is interpreted as a descriptive performance gain rather than a statistically significant improvement.