Ancient manuscripts represent valuable cultural heritage but are highly vulnerable to physical degradation and loss of public understanding. One such heritage is the Incung script manuscripts from Kerinci Regency, Indonesia, which remain underexplored in digital recognition research. This study proposes an optimized integration of image enhancement, template matching, and Convolutional Neural Network (CNN) methods to improve the accuracy and stability of Incung character recognition. Image enhancement is applied to improve contrast and legibility, template matching is used to capture structural character patterns, and CNN is employed as a complementary classifier to validate recognition results. Experimental results on segmented character images show that the proposed approach achieves a training accuracy of 93% and a validation accuracy of up to 100% with stable loss values, indicating effective learning under controlled conditions. Although performance decreases when applied to full manuscript images due to segmentation challenges and low contrast, the proposed method demonstrates strong potential for digital preservation of Incung manuscripts. While this study does not directly implement smart city services, the resulting digital manuscript data can support local wisdom–based smart city initiatives by enabling digital cultural documentation and access.
This review examines how organizations are integrating cloud computing with artificial intelligence (AI) to enhance their digital marketing capabilities. We scrutinised scholarly articles, industry white papers, and illustrative corporate case studies covering the 2018–2025 period, with the aim of mapping the synergistic effects these technologies exert on widely recognised marketing pain points: difficulty in creating personalized customer experiences, wasting resources, and slow response to customer needs. Our analysis reveals five principal advantages: better business flexibility, improved customer tageting, real-time data analysis, automated processes, and cost saving. Quantified evidence suggests that firms adopting the AI-cloud model can lower operating costs by 70%, lift sales rates by 50%, elevate customer satisfaction scores by 60%, and accelerate decision-making by 40%. However, the implementation landscape features severe constraints, including data privacy reported by 78% of companies, a perceived over-reliance on external technology partners by 65% of respondents, and lack of skilled workers by 72%. Based on the findings, the study issues concrete guidance to executives: diversify cloud vendor portfolios, engineer interpretative user interfaces for AI applications, and commit to a robust training agenda. By synthesising contemporary evidence of advantages and vulnerabilities associated with AI-cloud marketing architectures, the research closes a significant knowledge void and furnishes a structured lens through which the evolving digital marketing landscape of the AI-cloud epoch can be systematically apprehended
Diabetic Retinopathy (DR) is a microvascular complication of diabetes and a leading cause of preventable blindness. Manual diagnosis through fundus imagery is time-consuming and requires highly specialized expertise. This study proposes an automated DR detection system through a comparative multi-scenario framework that evaluates independent clinical and semantic descriptors using a Multi-Layer Perceptron (MLP) network. To prevent majority class bias, a random undersampling technique was applied to the MESSIDOR dataset, resulting in a balanced dataset of 700 fundus images (350 Grade 0 and 350 Grade 1 instances). The system implements a dual-stream preprocessing pipeline utilizing Ben Graham's method for illumination standardization and Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance clinical lesions. This study independently analyzes texture representations from the Gray Level Co-occurrence Matrix (GLCM) and deep semantic features from a pre-trained EfficientNetB3 model. To mitigate computational overhead and explicitly prevent data leakage, feature standardization and Principal Component Analysis (PCA) were strictly isolated and executed within the training phase of each fold during the 5-Fold Cross-Validation protocol. The evaluation demonstrated that the proposed comparative framework achieves robust computational efficiency alongside a promising diagnostic testing accuracy of 97.00%. While these empirical results indicate strong internal reliability with minimal prediction errors, further evaluation involving external validation datasets is recommended to fully substantiate its utility as a supportive clinical screening tool.
This study aims to evaluate the performance of the IndoBERT model in sentiment classification of user reviews for the DeepSeek application on the Google Play Store. The reviews were categorized into three sentiment classes: positive, neutral, and negative. The dataset was collected through web scraping of Indonesian-language reviews and processed using several preprocessing stages, including cleaning, stopword removal, and stemming. This study contributes by systematically comparing hyperparameter optimization methods using Grid Search and Random Search under two data split schemes (60:20:20 and 80:20). In addition, oversampling and Focal Loss techniques were implemented to address class imbalance and improve neutral class classification. Experimental results show that the best performance was achieved using Grid Search with an 80:20 data split, resulting in a testing accuracy of 80.40% and a macro F1-score of 70.85%. This configuration also produced a lower GAP value, indicating better model generalization and reduced overfitting. The findings demonstrate that appropriate hyperparameter optimization significantly improves IndoBERT performance for Indonesian sentiment analysis tasks
The measurement of regional welfare in Indonesia is still dominated by economic indicators and therefore does not fully capture broader social conditions. This study aims to model and analyze multidimensional social welfare across provinces in Indonesia using a machine learning approach. The analysis is based on official data from Statistics Indonesia (BPS), covering five key dimensions: health, economy, education, infrastructure, and employment. The data were transformed using Benefit–Cost normalization and analyzed using the K-Means clustering algorithm. The results reveal four distinct welfare clusters with varying characteristics. The highest welfare cluster achieves an average index value of 0.6143, while the medium and lower-middle clusters record values of 0.4673 and 0.4210, respectively. The lowest welfare cluster has an index value of 0.3468, indicating limitations in health, economic, and employment dimensions. These findings highlight significant disparities in regional welfare conditions and demonstrate the importance of a multidimensional analytical approach. This study contributes by integrating multidimensional welfare indicators with a machine learning clustering method to provide a more objective and data-driven classification of regional welfare.
The challenge faced by farmers today is that it is still difficult to determine the types of vegetable crops that are suitable for cultivation according to environmental conditions, namely climate, soil type, and water requirements. Inaccuracies in choosing crop types are often caused by limited knowledge and lack of access to information related to the influence of environmental factors on the success of cultivation. This constraint makes it difficult for farmers to identify the characteristics of the selection of commodity types of vegetable crops that should be cultivated, this has an impact on suboptimal yields and limited sustainability in agriculture. Decision Support System (DSS) technology can help overcome this problem with its ability to determine the suitability of vegetable cultivation. The Multifactor Evaluation Process (MFEP) method is used in this study to assess and rank alternatives based on criteria that have been given weight values based on the parameters of temperature, sunlight, rainfall, humidity, slope, soil pH, and altitude. The results of the application of the MFEP method show that the system can provide recommendations based on validation testing accuracy values reaching 81.8%, precision 100%, recall 71.4%, specificity 100%, and f1-score 83.3%. These results show that DSS using the MFEP method can provide relevant recommendations by assisting farmers in choosing the type of vegetable crops that best suit the environmental conditions in their area. The benefits of this research can facilitate farmers in recommending types of vegetable crops that are suitable for cultivation to increase productivity and quality of crops.
Machine learning has the potential to support hematologists in classifying leukemia by identifying abnormal chromosomes and specific gene markers. One effective technique for feature selection is Variable-Length Particle Swarm Optimization (VLPSO), where its performance depends heavily on parameter control, specifically the inertia weight (w) and acceleration factors (c), which regulate the search process. In previous VLPSO, static types of parameter control were applied to the Factor, and time-varying types were used by the Factor. Although its results showed good performance in VLPSO, there was no separation in the treatment of training data and test data, leaving a gap in understanding their impacts for real-world applications. This study explores how different parameter control strategies (static, time-varying, and adaptive) affect the performance of VLPSO with two comparison adaptive parameter control approaches, Adaptive 1 and Adaptive 2, in the VLPSO framework, each designed to dynamically adjust the control parameters w and c in different ways. The 10-fold cross-validation shows that VLPSO with an Adaptive one-parameter setting achieves better generalization with low train-test differences, especially in Decision Tree and Naïve Bayes classifiers, though with higher variability. Adaptive 2-parameter setting of VLPSO offers more consistent results with narrower variability across different settings. Static methods are the least reliable, while time-varying controls show moderate but unstable performance. Adaptive parameter tuning is recommended to improve VLPSO's stability, flexibility, and classification accuracy in biomedical applications. The results provide recommendations for parameter settings using an adaptive approach that has been proven to enhance the performance of VLPSO
An integrated decision support framework for evaluating and ranking public complaint applications operated by city governments in West Java Province, Indonesia, is presented in this study. The Analytical Hierarchy Process is employed to derive expert-based weights for the ISO/IEC 25010 software quality criteria, while the Simple Additive Weighting method is applied to generate objective rankings of the applications. A research gap in the e-government domain is addressed, as limited studies have systematically implemented a multi-criteria decision-making approach that integrates expert consensus with quantitative evaluation for prioritizing digital public services. Data were collected through expert panel assessments and structured questionnaires administered across nine city-level complaint systems. Significant variations in software quality among the evaluated applications were identified, with Depok and Bandung achieving the highest scores of 0.8053 and 0.7593, respectively, while Bekasi and Cimahi obtained the lowest scores. Specific recommendations for improvement are provided based on the eight quality dimensions defined in the ISO/IEC 25010 model, and a replicable evaluation model for government stakeholders is proposed. A theoretical contribution is established through the extension of multi-criteria decision-making applications in evaluating e-government software quality using the comprehensive ISO/IEC 25010 framework. A practical contribution is delivered through the provision of a structured benchmarking tool that enables local governments to prioritize improvements in digital complaint systems based on measurable software quality indicators
Accurate extraction of Population, Intervention, Comparison, and Outcome (PICO) elements from clinical texts is essential for supporting evidence-based medicine, particularly in cardiology where clinical data complexity presents significant challenges. This study investigates the comparative effectiveness of three contextual embedding models—BioBERT, PubMedBERT, and SciBERT—integrated with a Bidirectional Long Short-Term Memory (BiLSTM) architecture for multi-label PICO element classification on heart disease datasets. The experimental framework involved pre-processing clinical sentences, transforming them into contextual embeddings, and classifying PICO elements using BiLSTM-based sequence modeling. Evaluation was conducted using five key metrics: accuracy, precision, recall, F1-score, and hamming loss, supplemented by confusion matrix analysis for each PICO element. Results demonstrate that the BioBERT-BiLSTM model achieved superior performance, with an accuracy of 73.89%, F1-score of 78.54%, precision of 81.60%, and recall of 76.64%. PubMedBERT-BiLSTM exhibited the highest precision (84.12%) but lower recall, while SciBERT-BiLSTM produced slightly inferior results overall. These findings confirm the importance of using domain-specific embeddings, particularly models pre-trained on biomedical corpora, to improve classification accuracy in specialized clinical text tasks. This study concludes that the BioBERT-BiLSTM combination offers a reliable approach for automated PICO element extraction in the cardiology domain, contributing to the development of more accurate and efficient clinical decision-support systems
The use of persuasive language is one of the defining features of native advertisements. Therefore, detecting persuasive content in news is essential, since native ads often appear disguised as legitimate news articles, it is crucial to identify and filter such content to maintain objectivity and improve the user experience. This study aims to detect news with persuasive content i.e. persuasive news in English language using a natural language processing (NLP) approach. The proposed method incorporates text summarization methods, pre-trained word embeddings, and deep learning models. An additional Conv1D layer has been added to improve the model’s performance. The model were trained on an Indonesian news dataset translated into English using Google Translate API. Experimental results show that our proposed RoBERTa–BiLSTM-Conv1D model, outperformed other models, achieving 92% accuracy in identifying persuasive news in English. These findings indicate that the persuasive content detection model can be used for application in mainstream media environments to detect native ads in English language. In the future, the model can incorporate Indonesian and English news as training data to develop a cross-lingual native ads detection model
Access to digital libraries and streaming services is often limited in regions with no internet connectivity, commonly referred to as blank spot areas. This paper presents the design and implementation of an offline server system to provide digital library and multimedia streaming services in several villages on Bengkalis Island, Indonesia. The system was installed in three villages and utilised a local server equipped with over 1,200 digital learning resources, distributed to users through a wireless local area network (Wi-Fi) without the need for internet access. During the three-month pilot period, 94% of users (n = 127) reported successful access to e-books and educational videos via their personal devices, with average access speeds reaching 40–48 Mbps. Content updates were performed every four weeks via temporary internet connections. User satisfaction was high, with 91% of respondents stating that the system helped them access up-to-date learning materials and improved their learning motivation. These results demonstrate that the offline server solution effectively bridges the digital divide, enabling equitable access to information, educational resources, and multimedia content for underserved communities in remote areas.
This systematic literature review analyzes machine learning approaches for mobile phone price prediction based on device specifications through a comprehensive examination of 25 research studies from 2018 to 2024.The review reveals that ensemble methods, particularly Random Forest (achieving up to 97% accuracy) and Gradient Boosting (R² = 0.9829), consistently outperform individual algorithms across various datasets. Support Vector Machine models demonstrate superior classification performance with 96-97% accuracy, while neural networks show perfect best-performer ratios but remain underutilized (4.88% of implementations). The following keywords were used in this systematic review's extensive search strategy across IEEE Xplore, ACM Digital Library, ScienceDirect, and Google Scholar: ("mobile phone price prediction" OR "smartphone price prediction") AND ("machine learning" OR "artificial intelligence") AND ("specifications" OR "features") AND ("classification" OR "regression"). Strict inclusion/exclusion criteria were used to select 25 studies from an initial pool of 45 studies, with an emphasis on empirical research with quantitative performance metrics published between 2018 and 2024. The study reveals RAM, internal memory, battery capacity, and processor specifications as the key determining features for mobile phone pricing. According to the study, the primary factors influencing mobile phone pricing are processor specifications, RAM, internal memory, and battery capacity. This review identifies critical research gaps, including insufficient neural network exploration, poor dataset reporting practices (52% of studies omit dataset sizes), and lack of real-time market dynamics integration. The findings provide evidence-based guidance for researchers, manufacturers, and consumers in selecting optimal prediction algorithms and understanding key price-determining features in the evolving smartphone market. Study limitations include geographic bias toward specific markets represented in available datasets, limited access to proprietary datasets, and a primary focus on specification-based features that exclude market sentiment analysis
Native advertising is often difficult to detect because it resembles regular news articles. One indicator is the absence of diverse information sources or the reliance on a single perspective. Therefore, it is necessary to employ an extraction technique capable of consolidating various forms of identical entity mentions. This study integrates an NER model based on XLNet+BiLSTM+CRF with identical entity classification using Levenshtein distance features and static and contextual vector representations. The results show an F1-score of 93.71% at the entity level and 92.84% for identical entity identification, along with a list of unique citation sources. These findings demonstrate that this unique list can be an additional feature in detecting native advertising, which often relies on a single source. With an average unique entity coverage of 97.40%, the proposed architecture can extract unique entities within news articles
Color blindness affects approximately 8% of the global population, emphasizing the importance of early detection. The conventional Ishihara Test had several limitations, such as paper degradation, low interactivity, and difficulties in maintaining children’s focus. This research introduces AR-VISION, an Augmented Reality-based Ishihara Test application designed for inclusivity and child-friendliness. Employing a Research and Development approach with the Model Development Life Cycle (MDLC), AR-VISION was developed and evaluated through six main phases. The technical evaluation showed Algorithm C achieved the best compromise between speed, accuracy, and memory usage. User testing with 35 elementary school students indicated a significant increase in accuracy (from 72% to 91%) and engagement (from 60 to 87). In conclusion, AR-VISION enhanced the precision, interactivity, and motivation of children in color blindness screening, supporting SDG 3, 4, and 9 and Asta Cita No.4, while demonstrating the transformative potential of AR in health, education, and interactive learning
This study proposes a Long Short-Term Memory (LSTM) model to predict the hatching success of sea turtle eggs in the Anambas Islands Marine Conservation Area, Indonesia. Leveraging nesting data (2022–2024) provided by LKKPN Pekanbaru and associated environmental variables, the model’s performance was assessed across various configurations of time steps (2, 5, 7, 30, and 45 days) and data splits (ranging from 60:40 to 90:10). The optimal configuration—7-day time step with a 60:40 train-test split—yielded RMSE = 17.90, MAE = 8.67, and R² = 0.34. Results revealed strong seasonal nesting trends and statistically significant interspecies differences in incubation periods (p < 0.05). While the model demonstrated high predictive accuracy for standard incubation durations (30–45 days), performance declined in extreme cases, highlighting the need for location-specific environmental data. This research illustrates the practical application of LSTM for ecological time series forecasting and provides a machine learning framework to support decision-making in ecotourism scheduling and marine conservation planning in island-based coastal ecosystems.
This study addresses the challenges of text classification in domain-specific Natural Language Processing (NLP) within the medical field, which differs significantly from general NLP due to the presence of complex medical jargon and informal language in clinical documents. The primary objective of this research is to develop and evaluate a cancer-related text classification model by integrating the Naïve Bayes algorithm with Laplacian smoothing and optimizing its alpha parameter using Ant Colony Optimization (ACO). Specifically, the study aims to determine whether ACO can effectively identify the optimal alpha value that enhances the classification performance of the Naïve Bayes model. Experimental results demonstrate that with an alpha value of 0.27, the proposed model achieves an accuracy of 81.05%. This indicates that the combination of ACO and Naïve Bayes significantly improves classification efficiency and accuracy. The findings contribute to more accurate interpretation of clinical cancer-related texts, supporting better-informed decision-making in medical contexts
This study aims to develop an automatic essay answer assessment system based on Natural Language Processing (NLP) to reduce the time and effort required for evaluation. The system uses Cosine Similarity and Manhattan Distance as evaluation metrics and implements two text embedding methods—Term Frequency-Inverse Document Frequency (TF-IDF) and Bag of Words (BoW)—to represent the user’s answer text. The methodology begins with text processing and pre-processing, followed by embedding and similarity calculation between the user’s answer and the reference text to generate an evaluation score categorized into three levels: good, sufficient, and poor. Based on Cohen’s Kappa analysis, the kappa value for Cosine Similarity reaches 0.78, indicating high agreement between the Cosine TF-IDF and Cosine BoW methods. In contrast, Manhattan Distance yields a kappa value of -0.05, indicating a discrepancy between the two Manhattan-based methods. The evaluation results suggest that Cosine Similarity is more suitable, whereas Manhattan Distance is not relevant for this task. At the modeling stage, the best classification models are Decision Tree and Random Forest, each achieving an accuracy of 96.67%. Although Random Forest demonstrates a higher AUC than Decision Tree, it requires a longer training time. Overall, the system is considered effective for assessing essay answers with both purpose and consistency, offering potential applications in the field of education
Inventory management in warehouses often experiences discrepancies in recording the number of cardboard boxes due to errors during the manual recording process. To overcome this problem, a cardboard detection method was developed using the Default Mask R-CNN model and a modified model using MobileNetV3 Small. The training data was obtained from a collection of cardboard photos which then went through an annotation stage. In the cReonfiguration stage, various anchor scales were applied to determine the bounding box parameters, while the training process used Stochastic Gradient Descent (SGD). The default model is trained with the initial Mask R-CNN settings, while the custom model modifies the backbone and Feature Pyramid Network (FPN) adjustments. The test results show that the custom model has higher efficiency with a parameter count of 20,857,704 and an average FPS of 10.92. However, the accuracy level of the custom model is lower than that of the default model
Libraries collect a lot of important transaction data, but they rarely use this information to improve how consumers interact with them. This work tries to bridge this gap by offering a novel use of machine learning to analyze and classify library patron behavior. Customers were categorized based on their age range, checkouts, and renewals using the KMeans clustering technique. Dimensionality reduction methods like PCA and t-SNE were used to visually clarify the generated patterns. Our research revealed three different user groups: Rare Borrowers, who typically make 5.4 checkouts and 2.0 renewals; Occasional Borrowers, who typically make 20.8 checkouts and 7.1 renewals; and Frequent Borrowers, who often make 50.3 checkouts and 15.4 renewals. The clustering model performed quite well, as evidenced by its Calinski-Harabasz Index of 320.12, Davies-Bouldin Index of 0.45, and Silhouette Score of 0.62. Beyond these metrics, the study’s novelty lies in its practical implications—offering libraries a data-driven framework to tailor services, improve user satisfaction, and optimize resource allocation. This study highlights the transformative potential of machine learning in library science offering a data-driven framework for libraries to personalize services, optimize book recommendations, and enhance outreach efforts based on patron behavior. By segmenting users, libraries can better allocate resources and improve user experience. Limitation of this study lies on the data bias which may affect generalizability due to demographic differences across libraries. Additionally, KMeans clustering assumes predefined clusters, which may not fully capture nuanced behaviors.
One of the most widely grown horticulture fruits in Indonesia is the banana. In addition to its various health benefits, bananas are a good source of carbohydrates and vitamins A, C, and E. There are a lot of different kinds of bananas in Indonesia, and occasionally people have trouble telling them apart. This study uses a Convolutional Neural Network (CNN), a Deep Learning technique, to categorize bananas. Four different types of bananas—Cavendish, Kepok, Raja, and Tanduk—were classified. Planning, analysis, creating a banana classification model with CNN, and assessing the outcomes are the four phases of the research process. Data preprocessing, CNN model creation, training, and testing procedures are the next steps in the categorization model design process, which starts with the collection of banana data using a smartphone camera. The optimal model was obtained with the accuracy value of 96%, the average precision and recall values of 97% and 96% respectively. It was found based on test results with multiple tuning parameters, including dataset partition, optimizer use, and epoch. This study offers novelty in terms of the use of a large banana image dataset, extensive exploration of CNN parameters, and the potential application of the model in applications for the horticultural industry. In addition, this study contributes to the development of image-based AI technology in agricultural product classification, which is still relatively underexplored in Indonesia
