Journal of Computing and Smart Ecosystems

Provincial Classification of Child Malnutrition in Indonesia: A K-Means Clustering Approach for Evidence-Based Stunting Intervention Prioritization

Inas Nilasari — 2026-06-30

Indonesia's stunting prevalence of 21.5% recorded in SSGI 2024 surpasses the WHO emergency threshold, yet provincial data are conventionally presented in tabular formats that obscure spatial risk patterns and constrain evidence-based resource allocation. This study applies hierarchical clustering to 35 Indonesian provinces using five child nutritional indicators that is stunting, underweight, severe wasting, wasting, and overweight drawn from SSGI 2024. Four linkage methods were evaluated across k = 2 to 7 using Silhouette Score, Dunn Index, Connectivity, and Cophenetic Correlation Coefficient. Ward.D2 at k = 2 yielded the optimal configuration, producing two substantively distinct clusters: Cluster 1 (18 provinces) with comparatively lower undernutrition burden (mean stunting 17.98%, mean wasting 5.72%), and Cluster 2 (17 provinces) with markedly elevated risk across all undernutrition indicators (mean stunting 26.60%, mean wasting 9.11%). Papua consistently emerged as a structurally outlying unit across all linkage methods. These findings provide an empirical basis for geographically differentiated nutritional policy, enabling priority deployment of emergency interventions in high-burden provinces while sustaining preventive programming elsewhere.

Comparative Evaluation of EfficientNet-B0 and MobileNetV3 for Multi-Class Fire Image Classification with Grad-CAM Interpretation

Muhammad Fabian Rizky Fatah — 2026-06-30

Fire detection using image-based computer vision is a promising alternative to conventional sensor systems, which are often limited by detection range and false alarms. This study presents a comparative evaluation of two lightweight deep learning architectures, EfficientNet-B0 and MobileNetV3 Small, for multi-class fire image classification across four categories: Urban Fire, Wild Fire, Non-Damage Building, and Non-Damage Wildlife. Both models were trained using transfer learning with ImageNet pre-trained weights on a curated dataset of 1,933 images with selective undersampling to address class imbalance. Grad-CAM (Gradient-weighted Class Activation Mapping) was applied to interpret model decisions and validate the learned visual features. EfficientNet-B0 achieved the highest overall accuracy of 93.8% with a macro F1-score of 0.925, while MobileNetV3 Small reached 91.0% validation accuracy with a macro F1-score of 0.900. Grad-CAM visualizations confirmed that EfficientNet-B0 develops more localized feature activations focused on flame-colored regions, while MobileNetV3 exhibited broader contextual activations. The results demonstrate that lightweight transfer learning models combined with Grad-CAM provide both high classification performance and meaningful interpretability for fire detection applications.

Mechanistic Interpretability of Transformer Attention Heads in Indonesian Fiscal Policy Discourse: Decoding Internal Representations and Latent Biases

Zia Ul Rehman Zafar — 2026-06-30

Transformer-based language models are increasingly used in economic text analysis, yet their internal reasoning remains insufficiently interpretable, particularly in multilingual policy contexts. This study examines the mechanistic interpretability of attention heads in analyzing Indonesian fiscal policy discourse using a bilingual dataset comprising 47 YouTube transcripts and 23,500 news articles collected between 2018 and 2025. A fine-tuned IndoBERT model is implemented for fiscal stance classification, achieving strong performance with accuracy of 0.91, precision of 0.89, recall of 0.88, and F1-score of 0.885, ensuring a reliable basis for interpretability analysis. To uncover internal model behavior, an integrated framework combining attention entropy, head importance scoring, and Bias Sensitivity Index (BSI) is applied. Results show a consistent decline in attention entropy across layers from 2.85 to 1.62, indicating a shift from distributed contextual processing to focus semantic representation. Head-level analysis reveals functional specialization, with key attention heads reaching importance scores up to 0.25. Ablation experiments further demonstrate measurable prediction shifts, reflected in elevated BSI values, confirming sensitivity to specific heads associated with fiscal risk terms. These findings indicate that transformer models encode structured economic reasoning while exhibiting latent bias, highlighting the need for interpretability-driven approaches in AI-based policy analysis.

MoodyMap: An Interactive Map Application for Saving User Stories and Moods Based on Location

Naimatul Husna — 2026-09-03

MoodyMap is an interactive map application designed to allow users to record and revisit personal stories and moods based on geographical locations. The application was developed using the Software Development Life Cycle (SDLC) with an Agile approach, utilizing Laravel and Bootstrap as development frameworks and Leaflet.js for interactive map visualization. MoodyMap provides several main features, including user registration and authentication, adding location-based memories with automatic coordinates, mood representation using emojis, photo uploads, memory management, memory history, profile management, and password reset via email. Functional testing using black-box testing showed that the main features operated according to the expected results. The developed application demonstrates that integrating geographical information, personal narratives, and mood representation can provide a more contextual way for users to document and revisit their location-based memories.

Accuracy-Efficiency Benchmarking of Lightweight Machine Learning Models for Building Heating and Cooling Load Prediction

Nuki Pujiani Yosephine — 2026-06-30

Early-stage estimation of heating and cooling loads supports energy-efficient building design, but complex predictive models may impose unnecessary computational costs for small tabular datasets. This study benchmarks four lightweight regression models, such as Linear Regression, Ridge Regression, Random Forest, and Gradient Boosting, using the UCI Energy Efficiency dataset containing 768 simulated building configurations, eight design variables, and two continuous targets. Model accuracy was evaluated with shuffled 10-fold cross-validation using mean absolute error (MAE), root mean squared error (RMSE), and the coefficient of determination (R²). Computational efficiency was assessed through training time, prediction latency, and serialized model size, while permutation importance was used for interpretation. Random Forest achieved the lowest heating-load RMSE (0.4624) and an R² of 0.9978; however, its difference from Gradient Boosting was not statistically significant. Gradient Boosting was approximately 96 times smaller and 20 times faster at inference. For cooling load, Gradient Boosting achieved the lowest RMSE (1.4903) and an R² of 0.9751, significantly outperforming Random Forest in fold-level RMSE. Relative compactness was the most influential heating-load feature, whereas overall height dominated cooling-load prediction. The results indicate that Gradient Boosting offers the strongest overall accuracy-efficiency trade-off for lightweight smart-building prediction systems.

Comparative Sentiment and Complaint Pattern Analysis of Indonesian E-Wallet Application Reviews Using IndoBERT

Firstiawan Fadhil Thobibuddin — 2026-06-30

The rapid adoption of e-wallet services in Indonesia has generated a large volume of user-generated reviews that provide valuable information about service quality, user satisfaction, and recurring technical problems. However, previous studies have commonly examined individual applications or emphasized classification performance, while cross-platform comparisons using a consistent Indonesian language model remain limited. This study conducts a comparative sentiment and exploratory complaint-term analysis of five widely used Indonesian e-wallet applications—DANA, OVO, GoPay, ShopeePay, and LinkAja—using IndoBERT. A total of 5,000 Google Play Store reviews were initially collected, of which 4,669 reviews remained after filtering and preprocessing. Ratings of 1–2 were used as negative proxy labels, whereas ratings of 4–5 were used as positive proxy labels, while three-star reviews were excluded. The dataset was divided into 80% training and 20% testing data using stratified sampling. The indobenchmark/indobert-base-p1 model was fine-tuned for three epochs using a learning rate of 2×10⁻⁵ and a batch size of 16. The best model achieved 84.36% accuracy and a positive-class F1-score of 73.45%. At the platform level, DANA recorded the highest proportion of positive reviews at 51.84%, whereas OVO showed the highest negative proportion at 84.77%. Exploratory complaint-term analysis identified recurring issues involving failed transactions, account balances, login and verification problems, application errors, and customer service. The findings demonstrate the usefulness of IndoBERT for analyzing Indonesian e-wallet reviews and provide a cross-platform perspective for identifying service improvement priorities in Indonesia's digital payment ecosystem.

Development of an Interactive Website-Based Learning Media for C++ Operator Topics in Informatics Education

Yoseph Satria Praka — 2026-06-30

Computational thinking has become an essential competency in informatics education, yet many senior high school students still encounter difficulties in understanding fundamental programming concepts, particularly C++ operators. This study aims to develop and evaluate an interactive website-based learning media for C++ operator topics to support the improvement of students’ computational thinking skills. The learning media integrates structured instructional materials, interactive programming games, and formative quizzes within a single web platform to promote active and self-paced learning. The study employed a one-group pretest–posttest design involving ten students from grades 10 to 12 who were members of the IT Club at a partner senior high school. Students completed pretests and posttests to measure their computational thinking performance before and after using the learning media. The effectiveness of the developed media was evaluated using the Normalized Gain (N-Gain) analysis to determine the magnitude of learning improvement. The results show that the average pretest score increased from 21.00 to 51.00 in the posttest, with an average N-Gain value of 0.37, which falls into the medium improvement category. These findings indicate that the developed interactive website-based learning media effectively supports students’ understanding of C++ operators and contributes to the development of computational thinking skills through engaging and interactive learning activities.

Smart Ecosystem Performance Optimization Using Federated Learning for IoT-Based Real-Time Data Processing

Muhammad Distian Andi Hermawan — 2026-06-30

The rapid expansion of smart ecosystems driven by Internet of Things (IoT) technologies requires real-time data processing that is secure, efficient, and capable of handling large-scale heterogeneous data streams. Traditional centralized learning approaches often struggle with latency, bandwidth constraints, and privacy risks due to the continuous transfer of raw data. This study aims to optimize smart ecosystem performance by integrating Federated Learning (FL) as a distributed intelligence framework for real-time IoT data processing. The proposed model allows multiple IoT nodes to collaboratively train a global model without sharing raw data, thereby enhancing privacy protection and reducing network load. Experimental results demonstrate that FL provides higher accuracy, better scalability, and improved resilience compared to centralized systems, especially in environments with diverse sensor distributions. These findings confirm that Federated Learning is a highly promising approach for strengthening smart ecosystem reliability, responsiveness, and data security. This research contributes to developing future-ready IoT architectures capable of supporting intelligent, decentralized, and privacy-preserving environments.