Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control

AIoT-Enabled Automatic Waste Sorting System with Real-Time WhatsApp Notifications

2025-12-17T07:15:35+00:00

The waste management crisis, particularly in educational institutions, requires innovative solutions that combine artificial intelligence and automation. This research develops and evaluates an automated waste sorting system based on Artificial Intelligence of Things (AIoT) integrated with WhatsApp notifications. The system utilizes the EfficientNet-B0 deep learning model optimized with transfer learning and runs on a Raspberry Pi 4 edge device to classify waste into five categories: plastic, paper, metal, glass, and organic in real time. Classification results are translated into physical actions by a servo actuator mechanism, while ultrasonic sensors monitor trash bin capacity. The real-time notification system via WhatsApp API sends alerts to administrators. A 30-day evaluation on campus showed that the system achieved 92.3% classification accuracy with an inference latency of 1.8 seconds. The mechanical system successfully sorted waste with a 94.5% success rate, and WhatsApp notifications had a 99.1% delivery rate, with an average administrator response time of 8.2 minutes during operational hours. A comparative analysis demonstrated that this system increased sorting efficiency by 87% and reduced operational costs by 45% compared to manual waste sorting methods. These findings conclude that the proposed integration of edge AI, mechanics, and WhatsApp notifications creates a smart waste management solution that is not only effective and real-time but also practical, economical, and sustainable for wider implementation.

A Memory-Efficient and Gradient-Stable Lightweight ANFIS for Real-Time Humidity Prediction in Precision Agriculture

2026-02-08T09:05:46+00:00

Precision agriculture demands artificial intelligence solutions that are both accurate and deployable on resource-constrained hardware, yet conventional machine learning models require excessive memory while traditional ANFIS architectures suffer from training instability. This study developed a memory-efficient and gradient-stable lightweight Adaptive Neuro-Fuzzy Inference System (ANFIS) for real-time humidity prediction on microcontroller-class devices. The proposed architecture strategically reduced the rule base from 27 to only 4 interpretable fuzzy rules and limited membership functions to two per input, achieving an 85.2% reduction in learnable parameters. A gradient-stable training mechanism was introduced, combining physics-informed parameter initialization with adaptive gradient clipping to prevent gradient explosion. The model was trained and validated using 31,474 real-world greenhouse samples collected over 218 days, with 80% allocated for training and 20% for temporal testing. Experimental results demonstrated that the gradient-stable architecture successfully converged from a catastrophic R² of -64.08 to 0.9148, with a root mean square error of 1.32% and mean absolute error of 1.05%. The model required only 0.211 KB of memory, representing a 99.9% reduction compared to baseline Random Forest models, while achieving inference time of 8.2 milliseconds on Arduino UNO. The system was successfully deployed on three independent hardware modules, maintaining consistent performance with average RMSE of 1.99% over 168 hours of continuous operation. This study concludes that strategic simplification and stability-aware training enable interpretable neuro-fuzzy systems to operate effectively on ultra-low-resource devices, bridging the gap between predictive accuracy and hardware feasibility in embedded agricultural IoT applications.

A Data-Driven Framework Integrating Clustering and Classification for Fair Tuition Grouping (UKT) Prediction

2025-12-17T06:53:56+00:00

This study aims to identify the most effective combination of feature selection techniques and classification algorithms for predicting student tuition groups (Uang Kuliah Tunggal, UKT) based on pre-admission data. Three feature selection methods Exploratory Factor Analysis (EFA), Recursive Feature Elimination (RFE), and Random Forest Feature Importance (RFFI) were employed and combined with five supervised learning models: Decision Tree, Random Forest, Support Vector Machine (SVM) with RBF kernel, Naïve Bayes, and K-Nearest Neighbor (KNN). The results demonstrate that the EFA–SVM (RBF) combination achieved the best performance, with an average accuracy exceeding 98%, outperforming other models across most faculties. EFA also yielded the highest Silhouette Score (0.2933), indicating a more stable and distinct cluster structure compared to RFE (0.2564) and RFFI (0.2575). These findings highlight the critical role of appropriate feature selection in improving classification accuracy and model generalization, particularly when emphasizing socioeconomic variables such as parental income, land area, housing conditions, and basic family facilities. The integration of factor-based dimensionality reduction with non-linear classification algorithms proved effective in developing a more transparent and equitable UKT prediction model. This research contributes to the advancement of data-driven decision support systems in higher education and provides a foundation for future automation in tuition group determination processes.

Dota 2 Hero Buff And Nerf Predictions Based On Professional Match Data Using Random Forest

2026-02-08T09:01:32+00:00

Balancing updates (buffs and nerfs) are critical in Multiplayer Online Battle Arena games because small parameter changes can shift the competitive metagame and reduce hero diversity. This study proposed a data-driven pipeline to classify each Dota 2 hero as overpowered, underpowered, or balanced from professional match telemetry and to translate these classes into balance recommendations (nerf, buff, or balance). Most prior Dota 2 studies focus on match outcome or micro-event prediction and do not evaluate hero-centric balance recommendations against official patch actions across patch transitions. To address this gap, this work contributes a patch-to-patch external validation protocol that compares recommendations from patch t with developer actions in patch t+1 using patch notes. Professional match records were collected from public sources and aggregated per hero and per patch into combat, economy, and impact features (e.g., kills, deaths, assists, gold per minute, experience per minute, damage dealt, tower damage, and healing). Labels were derived from win-rate and pick-rate distributions using statistical control limits (μ ± kσ, k = 0.3) to ensure transparent and repeatable labeling. A Random Forest classifier was trained using grid-searched hyperparameters and evaluated using stratified 6-fold cross-validation with macro-averaged F1 to address class imbalance. Internal evaluation achieved 0.94 accuracy and 0.84 macro-F1. For external validation, recommendations from patch t were compared with official balance actions in patch t+1 across six consecutive transitions; accuracy ranged from 0.436 to 0.672 (mean 0.559), with the best result on 7.39b to 7.39c (84/125). These results indicated that professional telemetry could support interpretable balance monitoring and provide early signals for buff/nerf candidate review

Predicting Social Media Post Engagement and Virality Using Graph Neural Network Approaches and Content-Based Features

2026-02-08T08:53:16+00:00

Social media teams increasingly rely on early signals to prioritize content, yet forecasting engagement and identifying viral posts remain difficult under temporal drift and heavy-tailed interaction counts. This study evaluated Graph Neural Network (GNN) approaches for predicting post engagement and virality from pre-posting content-based and contextual features. The Social Media Engagement Report dataset, which contained 100,000 posts across Twitter, LinkedIn, Facebook, and Instagram spanning March 2021–March 2024, was used. Post-release variables (impressions, reach, engagement rate) were excluded to prevent leakage. A homogeneous post–post graph was constructed using k-nearest-neighbor similarity in an embedding space and exact-match links on low-cardinality context. Ridge/Logistic Regression, Random Forest, and XGBoost as the baselines were compared against GraphSAGE and GAT under a chronological train, validation, and test split. Regression used MAE, RMSE, and R²^, while virality classification used ROC-AUC, PR-AUC, and Precision at the top 1% ranked posts. GraphSAGE yielded the strongest virality screening, achieving ROC-AUC = 0.66, PR-AUC = 0.54–0.56, and Precision@1% up to 0.75, substantially above non-graph baselines. For regression, GAT produced the lowest errors despite a negative R², indicating limited explained variance. Overall, similarity-graph GNNs are most effective for early virality identification, whereas exact count prediction remains challenging in a strictly pre-posting, time-aware setting.

Optimization of Retargeting Motion Capture for Remo Dance Using Fuzzy Logic

2026-01-16T21:50:45+00:00

Retargeting motion capture for traditional dance animation faces challenges in maintaining biomechanical accuracy while preserving cultural expressiveness, especially when human motion data are transferred to character models with different skeletal structures. This research aims to optimize the retargeting of East Java Remo Dance through an adaptive artificial intelligence-based evaluation approach. The Remo dance movement was recorded using a multi-camera optical motion capture system and retargeted to two types of 3D characters: realistic and stylized. The evaluation was conducted using quantitative metrics (Mean Squared Error, Structural Similarity Index, Dynamic Time Warping, and Kalman Filtering) as well as a qualitative approach through Laban Movement Analysis. Subsequently, Mamdani fuzzy logic was integrated to synthesize all these parameters into the Fuzzy Retargeting Quality Score (FRQS). The results showed that the realistic character had higher movement accuracy (MSE = 0.0032; SSI = 0.89; DTW = 0.92) and obtained an FRQS value of 86.4 (very optimal category), whereas the stylized character obtained an FRQS of 71.2 (moderately optimal), reflecting a compromise between movement precision and visual appeal. The integration of fuzzy logic allows for more contextual and human-centric retargeting evaluation, as well as strengthening the dual-model approach to the preservation and education of traditional dance based on digital animation.

Kalman Filter Based RSS Preprocessing for Cryptographic Key Generation in Zero Knowledge Feige Fiat Shamir Authentication

2026-03-10T02:56:51+00:00

Secure authentication in wireless communication environments required mechanisms that were capable of verifying identity without exposing confidential information. Zero-Knowledge Authentication addressed this challenge by enabling interactive identity verification without revealing secret credentials however, its performance strongly depended on the reliability of the cryptographic key generation process. This study investigated the use of Received Signal Strength as a source for cryptographic key generation and addressed the instability caused by noise and signal fluctuation in wireless channels. A preprocessing approach based on the Kalman Filter was proposed to improve the quality of Received Signal Strength measurements prior to key generation. The Kalman Filter was applied to reduce noise and enhance signal reciprocity between communicating nodes, ensuring that both parties generated identical cryptographic keys. The filtered signal values were then utilized to support the Zero Knowledge Feige-Fiat-Shamir authentication mechanism by replacing the conventional communication channel with keys derived from the preprocessed signal measurements.The performance of the proposed approach was evaluated through key consistency, entropy level, and bit mismatch rate between legitimate nodes. The experimental results showed that Kalman Filter–based preprocessing improved the stability of Received Signal Strength measurements and significantly increased the consistency of generated keys compared to unfiltered approaches. Consequently, the authentication success rate was enhanced while maintaining the confidentiality properties of Zero-Knowledge Authentication. These findings demonstrated that Kalman Filter assisted preprocessing effectively strengthened the security and reliability of cryptographic key generation for wireless Zero-Knowledge authentication systems.

Automated breast cancer cell counting: comparing multi-class segmentation and two-stage classification strategies

2026-03-13T03:53:10+00:00

The manual interpretation of Hematoxylin and Eosin (H&E) histopathology images for breast cancer diagnosis is hindered by time limitations and observer bias. This research seeks to create an automated system using Deep Learning for cell detection and classification, evaluating two key approaches: Multi-class Segmentation (single-stage) and Segmentation followed by Classification (two-stage). U-Net architecture was employed for segmentation, while MobileNetV2 and VGG16 were used for classification. The models were tested on the public IHC4BC dataset and primary data from Airlangga University Hospital (RSUA). The study also evaluated the impact of Resizing versus Tiling data processing strategies. Experimental results showed that while MobileNetV2 and VGG16 classification models achieved a high testing accuracy of 98.80%, the two-stage integrated system revealed a high counting error with a Mean Absolute Error (MAE) of 119.87 for positive cells, primarily due to under-segmentation of overlapping cells. In contrast, the Multi-class Segmentation approach utilizing the Tiling strategy demonstrated superior performance. This model effectively preserved spatial resolution and distinguished cell types simultaneously, achieving the lowest positive cell MAE of 18.46 and a negative cell MAE of 1.66. This study concluded that multi-class segmentation with a Tiling strategy was the most effective and accurate approach for automated cell counting in histopathology images.

Comparison of Word2Vec and GloVe performance in Bi-LSTM models for Indonesian news classification

2025-12-25T05:09:45+00:00

The explosion in the volume of textual data from digital news presents challenges in classifying content automatically and efficiently. For the task of classifying Indonesian-language news, this study aims to compare the performance of several word embeddings specifically Word2Vec using CBOW and Skip-Gram architectures and GloVe when applied to a Bidirectional Long Short-Term Memory (Bi-LSTM) model. This study uses a dataset consisting of 6,715 news articles from the Indonesian news portal that have undergone pre-processing, divided into five categories. The model was trained using 80% of the training data with K-Fold Cross Validation (K=5), while the remaining 20% of the data was used for testing. The experimental findings indicate that the Bi-LSTM model, when combined with CBOW embedding, yielded the best performance, achieving 95.16% accuracy and a 95.15% F1-Score. The Skip-Gram model followed with solid performance, achieving an accuracy of 93.30% and the fastest computation time. Conversely, the model that used pre-trained GloVe embedding delivered the poorest performance, achieving 88.98% accuracy. This result suggests that training embeddings on a specific domain is more effective at capturing local context. The conclusion of this study confirms that selecting a word embedding method specifically trained on local datasets is also an important step in achieving optimal accuracy in Indonesian news text classification.

Accuracy Comparison of Multivariate Newton-Raphson, Newton-Kantorovich, and Levenberg–Marquardt Methods for Solving Nonlinear Systems Using Numerical Simulation

2025-12-23T13:06:25+00:00

Multivariable nonlinear equation systems often appear in engineering, physics, economics, and artificial intelligence modeling, but often do not have closed analytical solutions. Therefore, accurate, efficient, and stable numerical methods are needed. This study aims to comparatively evaluate three iterative methods, namely Multivariate Newton-Raphson, Newton-Kantorovich, and Levenberg–Marquardt, in solving identical high-complexity multivariable nonlinear systems. Simulations were performed using MATLAB with an error tolerance of 0.001 and a maximum iteration limit of 100. The test system consisted of a combination of trigonometric, exponential, and polynomial functions, resulting in nonlinear interactions that were challenging for each method. The simulation results show that Levenberg–Marquardt excelled with only 6 iterations and a final error of 3.246 × 10⁻¹⁰, indicating high stability and efficiency, followed by Multivariate Newton-Raphson with 13 iterations and an error of 4.606 × 10⁻⁹, while Newton-Kantorovich requires 27 iterations with an error of 5.770 × 10⁻⁷, reflecting slower semi-local corrections.Three-dimensional visualization shows the intersection point of the surface as a solution, providing an intuitive understanding of the iteration trajectory characteristics of each method. The novelty of this research lies in the integrated numerical simulation framework that allows direct quantitative comparison of the three methods on identical systems with the same initial conditions, tolerance, and iteration limits. These findings provide important empirical references for selecting efficient and stable iterative methods for multivariable nonlinear systems, as well as practical guidance for numerical applications in engineering, physics, and scientific computing.

Electrocardiogram Signal Analysis Based on Discrete Wavelet Transform with Machine Learning Method in Autistic Children

2026-04-05T16:42:40+00:00

ASD is a neurodevelopmental disorder that affects a child's ability to manage emotions, interact socially, and respond to the environment. The main challenge in monitoring children's physiological condition is the limited availability of objective observation methods that rely heavily on health professionals. One potential objective approach is to analyze the ECG signal. However, ECG signals in children with ASD generally have high levels of noise due to body movements during recording, making manual analysis and conventional methods difficult. This study aims to develop a classification system for the physiological condition of children with ASD based on ECG signals, specifically to distinguish between quiet and active states. The dataset consists of 1000 from each of the two active classes and 1000 from the quiet class. ECG signals were processed using DWT for filtering, and then classified using three machine learning algorithms: SVM, RF, and AdaBoost. The performance of each model was evaluated using accuracy, precision, recall, and F1-score metrics. The evaluation results showed that Random Forest provided the best performance, with an accuracy value of 93%. Meanwhile, SVM achieved an accuracy of 91.25%, while AdaBoost showed slightly lower performance at 90.00%. Based on these results, Random Forest was selected as the most optimal model and integrated into a web-based system using Streamlit. This study demonstrates that the combination of DWT and Random Forest is effective for classifying the physiological conditions of autistic children and has the potential to serve as an objective tool for monitoring them.

Regularization Techniques to Improve the Stability and Accuracy of MLC Algorithm

2025-12-17T07:04:33+00:00

Maximum Likelihood Classification (MLC) is a classification algorithm that has important applications in the fields of image processing and remote sensing. No use of MLC was found in other fields. MLC assumes that data comes from a certain probability distribution (for example, a normal distribution), which may be too simple to describe complex data or have a non-normal distribution. This can lead to poor performance in situations where distribution assumptions are not met. That is why in various literatures there is no use of MLC for classification problems other than remote sensing. We propose a regularization technique to reduce distribution assumption errors in MLC called Regularization on maximum likelihood classification (RMLC). Regularization techniques are integrated into the covariance matrix, where regularization can make the data variance larger or smaller than the actual variance. This technique can also overcome singularities in the covariance matrix, non-Gaussian data, and data containing outliers. Experimental results on 13 public datasets show a significant increase in accuracy performance. The average accuracy increase reaches more than 11%, from 0.802 to 0.919, highlighting its potential for broader applicability and enhanced performance

A Gradient Boosting–Based Platform with Fuzzy Linguistic Representation for Cardiovascular Disease Risk Prediction

2026-02-08T09:04:14+00:00

Cardiovascular disease (CVD) is one of the most common causes of death around the world. In order to effectively prevent and manage CVD, early detection and prediction of risk are essential. This research introduces a healthcare platform based on CVD risk prediction using advanced machine learning (ML) methods. This platform is designed to provide accurate risk assessment by integrating the gradient boosting (GB) classifier method. Additionally, other ML models are used as comparison algorithms. Initially, this research used preprocessing techniques such as data normalization and data cleaning to tackle outliers in the dataset. Recursive feature elimination (RFE) feature selection approaches are utilized to find features that affect prediction performance, hence lowering the amount of data dimensions and enhancing model performance. Then, using metrics such as accuracy, precision, recall, and F1-score, each model’s performance is evaluated. The modeling results of the suggested approach are then used to create a digital health platform that predicts new input from users. Additionally, fuzzy logic is applied to transform data into linguistic variables to help users find simpler information. Using the proposed GB model and preprocessing method, the platform can make more accurate CVD risk predictions during data validation than other ML methods. When compared to other approaches with lower accuracy, the evaluation results demonstrate that the GB method can achieve the highest prediction accuracy of 94.30%.

Federated Learning and Deep Reinforcement Learning Synergy: Opportunities for Multi-Cloud Serverless Deployment

2026-02-08T08:59:33+00:00

The Development of distributed computing has enabled the use of multi-cloud and serverless computing, which are beneficial due to their flexibility, scalability, and cost efficiency. There are, of course, pertinent challenges associated with these computing paradigms, such as resource heterogeneity, cold-start latency, vendor lock-in, and privacy. Recent trends in Federated Learning (FL) and Deep Reinforcement Learning (DRL) hold promise in solving these issues. FL systems enable decentralised, privacy-preserving model training across heterogeneous systems, while DRL systems enable adaptive models for real-time decision-making to optimise system resources and improve performance. This Systematic Literature Review (SLR) covers the years 2020 to early 2026 and examines the intersection of FL and DRL in multi cloud serverless computing, following the PRISMA methodology. A primary analysis of 50 quality studies was undertaken to answer four privacy-related resource management questions. The results showed FL improves privacy and scalability using decentralised training. Consolidating the Federated DRL and Multi-Agent stacks enhances the system by achieving a better trade-off and optimization among latency, energy, and operational efficiency. However, a few gaps still exist, such as the absence of a more holistic framework, elusiveness in cross-system integration and collaboration, and a lack of concrete real-world applications. More work is needed to build a cohesive Federated Learning framework to improve sustainability and security in the multi-cloud, serverless systems of the future. This examination provides a solid foundation for the Development of innovative, privacy-preserving, and dynamic resource management in future cloud computing environments.

Federated Ensemble Learning with SHAP–LIME Interpretability for Smart Home Energy Prediction

2026-01-23T08:13:36+00:00

The increased adoption of IoT-based Smart Home systems in Indonesia has resulted in a growing volume of device-level energy data, opening up opportunities for the development of predictive models to support efficient household electricity consumption. However, challenges related to accuracy, interpretability, and data privacy remain a major concern, especially when data is distributed across multiple devices. This study evaluates the performance of four tree-based ensemble models, namely Random Forest, Gradient Boosting, XGBoost, and LightGBM, in centralized learning and federated learning scenarios using the Indonesia Smart Home Dataset. After undergoing feature preprocessing and refinement, including the removal of Sofa Pressure and Bed Pressure due to high noise, each model was trained and evaluated using MAE, MSE, and RMSE metrics. Federated learning was implemented through the Federated Averaging (FedAvg) algorithm to maintain data privacy without the need to transfer raw data between devices. The results show that LightGBM consistently provides the best performance in both scenarios and demonstrates resilience to data fragmentation and heterogeneity. Although there was a slight increase in error in federated learning, the error values remained within an acceptable range. SHAP and LIME analyses revealed that high-power devices such as air conditioners, water pumps, rice cookers, lights, and refrigerators had the greatest contribution.

Poultry Disease Classification Using EfficientNetV2-L and MobileNetV2 Based on Fecal Images

2026-01-16T21:29:31+00:00

Poultry diseases have a significant impact on livestock productivity; therefore, early detection is crucial to prevent infection spread. Deep learning approaches have recently shown promising results in improving disease classification accuracy. Convolutional Neural Network (CNN) models can identify poultry diseases through fecal images using automatic feature extraction. This study proposes poultry disease classification using two CNN architectures, EfficientNetV2-L and MobileNetV2. Each model was trained under three scenarios: baseline, class weights, and Focal Loss, using the Poultry Diseases Detection dataset from Kaggle consisting of four classes of chicken fecal images. The experimental results show that applying Focal Loss improves model performance compared to other scenarios. The EfficientNetV2-L model with Focal Loss achieved the highest accuracy of 99.51%, precision of 99.57%, recall of 99.51%, and F1-score of 99.52%. Meanwhile, MobileNetV2 performed reasonably well with faster training time. These findings indicate that combining Focal Loss with efficient CNN architectures enhances the classification of imbalanced datasets and has the potential to be implemented in real-time poultry disease detection systems

Scalable Multi-Agent Formation Control in RTS Games: A Virtual Anchor and Fluid-Based Allocation

2026-01-03T02:06:12+00:00

The control system for troop formation movement is a critical component in Real-Time Strategy (RTS) games, directly impacting gameplay quality and player experience. However, implementing these systems presents significant challenges, particularly in balancing rigid formation structure with pathfinding efficiency in dynamic environments containing complex obstacles. This study proposes an integrated framework for troop formation movement that synthesizes a virtual "Anchor" navigation paradigm with a "Fluid-Based Formation Position Allocation" algorithm. Unlike traditional leader-follower methods, the proposed system utilizes a virtual anchor to calculate global pathfinding via NavMesh, while constituent agents dynamically adjust their positions relative to this reference point. To mitigate trajectory conflicts during formation changes, the system employs a fluid-dynamics-inspired sorting strategy that deterministically maps agents to target slots using parallel processing. The architecture is optimized for real-time performance using the Unity Job System, allowing for the coordination of large-scale agent aggregates. Experimental validation was conducted through behavioral scenarios—including Tunnel, Split, and Crowd tests and stress tests involving up to 4,096 agents. The results demonstrate that the system successfully maintains formation integrity, executes autonomous regrouping after obstacle traversal, and ensures collision-free movement. Performance analysis indicates that the control logic remains computationally stable at scale, with the primary limitations shifting to graphical rendering overhead rather than algorithmic complexity.

A Comparative Study of Hybrid GARCH–HOLT–BPNN Models for Rainfall Forecasting Using a MATLAB-Based Intelligent Computing System

2026-01-03T01:20:16+00:00

Rainfall forecasting is a fundamental aspect of water resource management, hydrometeorological disaster mitigation, and agricultural planning, all of which are strongly influenced by climate variability. The complexity of rainfall data, characterized by non-linear, non-stationary, and highly fluctuating patterns, necessitates the use of adaptive and accurate predictive approaches. This study aims to conduct a comparative analysis of five forecasting models, namely the Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model, Holt’s Exponential Smoothing, Backpropagation Neural Network (BPNN), the hybrid GARCH–Holt model, and the advanced hybrid GARCH–Holt–BPNN model, in order to identify the most effective method for monthly rainfall forecasting. Rainfall data for the period 2015–2024 were used for model training and testing. Model performance was evaluated using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). In addition, this study incorporates the development of a MATLAB-based Graphical User Interface (GUI) to facilitate interactive model implementation and visualization of forecasting results. The results indicate that the GARCH model excels in capturing data volatility, Holt’s Exponential Smoothing effectively follows short-term trends with stability, and BPNN is capable of modeling non-linear relationships despite its sensitivity to data variability. The hybrid GARCH–Holt model demonstrates improved accuracy compared to single models. Furthermore, the hybrid GARCH–Holt–BPNN model achieves the most optimal performance, with an accuracy approaching 99% and the lowest MAPE value of 1.13%, reflecting excellent generalization capability. These findings confirm that the integration of linear and non-linear methods within a hybrid framework significantly enhances rainfall forecasting accuracy and contributes to data-driven decision-making in the field of hydrometeorology.

Hate Speech Analysis Using IndoBERT in YouTube Comments on the 2024 Indonesian Presidential Debate Video

2025-12-23T13:10:29+00:00

A Hate speech in the digital political space during election campaigns has the potential to cause polarization and undermine the quality of public discussion. This study analyzes hate speech in YouTube comments related to the five stages of the 2024 Indonesian presidential debate. We used IndoBERT, a Transformer-based language model specifically trained in Indonesian, to classify comments into hate speech and non-hate speech categories. The dataset consists of 38,742 comments collected from official debate videos. The dataset was labeled using a combination of manual annotation (20%) and semi-supervised learning (80%) using a pseudo-labeling approach. Experimental results show that IndoBERT achieved an average accuracy of 89.7% and a macro F1-score of 0.89 across all stages. IndoBERT outperformed baseline models such as mBERT, SVM, and Random Forest. These findings suggest that IndoBERT is more effective in capturing the linguistic nuances and distinctive Indonesian political rhetoric than multilingual or classical models. This study contributes an Indonesian-language political dataset and a comprehensive evaluation of relevant hate speech detection models for further research. Keywords: hate speech, IndoBERT, 2024 presidential debate, semi-supervised learning.