Publication Search

The development of social media platforms like TikTok has created new spaces for digital economic activities, including the practive of thrifting, which has now become a trend among the public. However, government policies that block these activities have sparked various public reactions. This study aims to analyze public sentiment regarding the issue of thrifting bans on the TikTok platform using the Bidirectional Long Short-Term Memory (Bi-LSTM) method. This method was chosen because it can understand text context from both directions, allowing it to capture deeper semantic meaning. The dataset consist of 4,000 TikTok user comments collected through a crawling process. The research stages include data preprocessing, sentiment labeling, splitting training and test data, training the Bi-LSTM model, and evaluating performance using accuracy, precision, recall, and F1-score metrics. The research results show that the Bi-LSTM model achieved an accuracy of 86.15%, with stable classification performance and minimal error rate. These findings indicate that Bi-LSTM is effective for sentiment analysis of public opinions on Indonesian language social media, particularly on context specific policy issues. Further development can be carried out by adding pre-trained embeddings or attention mechanisms to improve the model’s performance.

Polusi udara partikulat halus (PM2,5) merupakan ancaman serius bagi kesehatan masyarakat di Kabupaten Brebes, Jawa Tengah. Faktor penyumbang utamanya adalah emisi kendaraan di jalur Pantura, aktivitas industri perikanan, serta konsentrasi tinggi selama musim kemarau (Juni–November). Tidak adanya model peramalan sub-jam yang akurat menghambat pengembangan sistem peringatan dini yang efektif. Penelitian ini mengembangkan dan mengevaluasi model deep learning berbasis Transformer untuk memprediksi konsentrasi PM2,5 dengan resolusi waktu 15 menit. Data yang digunakan berasal dari NASA GEOS-CF (band PM25_RH35_GCC) yang diakses melalui Google Earth Engine menggunakan API Python. Dataset mencakup periode 1 Januari hingga 22 November 2025, menghasilkan 7.813 observasi per jam, yang kemudian diinterpolasi linear menjadi 31.249 titik data dengan resolusi 15 menit. Arsitektur Transformer terdiri dari 3 lapis enkoder, 4 kepala perhatian multi-head, dimensi embedding 128, dimensi feed-forward 256, panjang sekuen 60 timestep, dan augmentasi fitur menggunakan rerata bergulir (*rolling mean*, jendela = 3) dan beda pertama (*first difference*). Pelatihan dilakukan dengan TensorFlow-Keras, pengoptimal Adam, penjadwal peluruhan kosinus (*cosine decay scheduler*), dan fungsi kerugian Huber. Pembagian data dilakukan secara kronologis: 70% pelatihan, 30% validasi. Evaluasi pada set uji independen (16 Agustus–21 November 2025, 9.357 observasi atau 97 hari 11 jam 15 menit) menghasilkan MAE 0,7691 µg/m³, RMSE 1,2052 µg/m³, R² 0,9945, dan *Explained Variance Score* 0,9948. Model ini mampu menggambarkan variasi diurnal dan anomali musiman secara akurat, jauh melampaui model LSTM dan GTWR konvensional. Penelitian ini memberikan kontribusi signifikan di bidang Teknologi Informasi melalui kerangka kerja pengolahan *big data* satelit untuk aplikasi lingkungan.

Indonesian Sign Language (BISINDO) serves as a primary communication medium for the deaf community; however, limited public understanding often creates barriers during daily interactions. This study aims to develop a real-time BISINDO word-level translation system using hand landmark extraction and temporal modeling with Long Short-Term Memory (LSTM). The system employs MediaPipe Hands to detect 21 hand landmarks per frame, which are then processed as sequential motion patterns to classify five BISINDO words: saya, terima kasih, maaf, nama, and kamu. A total of 250 gesture samples were recorded under controlled lighting conditions as the primary dataset. The processed sequences were used to train the LSTM model, which was subsequently integrated with an ESP32 microcontroller and a DFPlayer Mini module to produce direct audio output. Experimental results show that the model achieved an average accuracy of 86%, with precision and recall values ranging from 0.81 to 0.94. The confusion matrix analysis indicates that most gestures were correctly classified, although some errors occurred in gestures with similar initial motion trajectories. Integration testing demonstrated an average system latency of 3.8 seconds and an audio output success rate of 85%. These findings indicate that the proposed system is capable of translating BISINDO word-level gestures accurately, responsively, and consistently in real-time conditions. This study provides a strong foundation for the broader development of sign language translation systems, with potential enhancements in vocabulary expansion, multi-user datasets, and hardware optimization for deployment in real-world environments.

This study aims to develop a web-based Virtual Assistant to improve the efficiency of academic information services at SMA Negeri 1 Parittiga. The research was motivated by the delays and inaccuracies in information delivery caused by the manual system still used in the school. The system development was carried out using the Research and Development approach with the Waterfall model, which includes the stages of needs analysis, design, implementation, and evaluation. The main technologies used are Natural Language Processing (NLP) and the Long Short-Term Memory (LSTM) machine learning algorithm, which allow the assistant to understand and respond to user questions in natural language in a contextual way. The system architecture uses Flask as the backend, Vue.js as the frontend, and Laravel for administrative data management. The testing results show that the system has an accuracy level of 88.4% in providing correct answers and a user satisfaction level of 92%, surpassing the target success rate of 80%. These findings prove that integrating NLP and LSTM can enhance the system's ability to understand conversational context and speed up the distribution of academic information. The study concludes that a web-based Virtual Assistant is an effective solution for the digitalization of school information services and has the potential to support the implementation of artificial intelligence technology in secondary education in Indonesia.

Chronic diseases such as chronic kidney disease (CKD), diabetes, and heart disease remain major causes of mortality worldwide, highlighting the need for accurate and interpretable diagnostic models. However, conventional machine learning methods often face challenges of limited generalization, feature redundancy, and class imbalance in medical datasets. This study proposes an integrated classification framework that unifies three complementary feature paradigms: classical tabular attributes, deep latent features extracted through an unsupervised Long Short-Term Memory (LSTM) encoder, and quantum-inspired features derived from a five-qubit circuit implemented in PennyLane. These heterogeneous features are fused using a feature-wise attention mechanism combined with an AdaBoost classifier to dynamically weight feature contributions and enhance decision boundaries. Experiments were conducted on three benchmark medical datasets—CKD, early-stage diabetes, and heart disease—under both balanced and imbalanced configurations using stratified five-fold cross-validation. All preprocessing and feature extraction steps were carefully isolated within each fold to ensure fair evaluation. The proposed hybrid model consistently outperformed conventional and ensemble baselines, achieving peak accuracies of 99.75% (CKD), 96.73% (diabetes), and 91.40% (heart disease) with corresponding ROC AUCs up to 1.00. Ablation analyses confirmed that attention-based fusion substantially improved both accuracy and recall, particularly under imbalanced conditions, while SMOTE contributed minimally once feature-level optimization was applied. Overall, the attention-guided AdaBoost framework provides a robust and interpretable approach for clinical risk prediction, demonstrating that integrating diverse quantum, deep, and classical representations can significantly enhance feature discriminability and model reliability in structured medical data.

The integration of Internet of Things (IoT) with cloud computing has revolutionized healthcare systems, offering scalable and real-time patient monitoring. However, optimizing response times and energy consumption remains crucial for efficient healthcare delivery. This research evaluates various algorithmic approaches for workload migration and resource management within IoT cloud-based healthcare systems. The performance of the implemented algorithm in this research, Hybrid Dynamic Programming and Long Short-Term Memory (Hybrid DP+LSTM), was analyzed against other six key algorithms, namely Gradient Optimization with Back Propagation to Input (GOBI), Deep Reinforcement Learning (DRL), improved GOBI (GOBI2), Predictive Offloading for Network Devices (POND), Mixed Integer Linear Programming (MILP), and Genetic Algorithm (GA) based on their average response time and energy consumption. Hybrid DP+LSTM achieves the lowest response time (82.91ms) with an energy consumption of 2,835,048 joules per container. The outcome of the analysis showed that Hybrid DP+LSTM have significant response times improvement, with percentage increases of 89.3%, 79.0%, 83.8%, 97.0%, 99.8%, and 99.94% against GOBI, GOBI2, DRL, POND, MILP, and GA, respectively. In terms of energy consumption, Hybrid DP+LSTM outperforms other approaches, with GOBI2 (3,664,337 joules) consuming 29.3% more energy, DRL (2,973,238 joules) consuming 4.9% more, GOBI (4,463,010 joules) consuming 57.4% more, POND (3,310,966 joules) consuming 16.8% more, MILP (3,005,498 joules) consuming 6.0% more, and the GA (3,959,935 joules) consuming 39.7% more. The result of ablation of the Hybrid DP+LSTM model achieves a 47.05% improvement over DP-only (156.57ms) and a 70.64% improvement over LSTM-only (282.41ms) in response time. On the energy efficiency side, Hybrid DP+LSTM shows 22.80% improvement over LSTM-only (3,671,51 joules), but 7.34% underperformance compared to DP-only (2,640,93). These research findings indicate that the Hybrid DP+LSTM technique provides the best trade-off between response time and energy efficiency. Future research should further explore hybrid approaches to optimize these metrics in IoT cloud-based healthcare systems.

The proliferation of fake news across digital platforms has raised critical concerns about information reliability. A notable example is the viral rumour falsely claiming that the Nigerian Minister of the Federal Capital Territory, Nyesom Wike, had collapsed at an event and was rushed to an undisclosed hospital an entirely fabricated claim that caused public confusion. While both traditional machine learning and deep learning approaches have been explored for automated fake news detection, many existing models have been limited to topic-specific datasets and often suffer from overfitting, especially on smaller datasets like ISOT. This study addresses these challenges by proposing a standalone Bidirectional Long Short-Term Memory (BiLSTM) model for fake news classification using the ISOT dataset. Unlike multi-modal frameworks such as the MM-FND model by state-of-the-art model, which achieved 96.3% accuracy, the proposed BiLSTM model achieved superior results with 98.98% accuracy, 98.22% precision, 99.65% recall, and a 98.93% F1-score. The model demonstrated balanced classification across both fake and real news and exhibited strong generalization capabilities. However, training and validation performance plots revealed signs of overfitting after epoch 2, suggesting the need for regularization in future work. This study contributes to the growing body of research on fake news detection by showcasing the efficacy of a focused, sequential deep learning model over more complex architectures, offering a practical, scalable, and robust solution to misinformation detection

The increasing number of mental health disorders in various countries has created an urgent need for innovation in the diagnosis and treatment process. This problem not only impacts individuals' quality of life but also creates a significant social and economic burden. One solution that is beginning to be widely researched is the use of artificial intelligence (AI) in the field of mental health. This research used a literature review of various previous studies discussing the role, application, and impact of AI. The results of the review indicate that AI technology, particularly in the form of digital applications such as chatbots, has great potential to support the recovery process for patients with mental disorders. AI-based chatbots can provide responsive, two-way interactions, so users feel heard and receive initial emotional support. One technical approach used is Natural Language Processing (NLP), which enables the system to understand natural human language. Simultaneously, Long Short-Term Memory (LSTM) algorithms are used to analyze language patterns and detect symptoms of depression more accurately. Various studies have reported that the application of NLP and LSTM can improve the reliability of diagnoses and provide responses tailored to user needs. Furthermore, AI can provide personalized recommendations, tailor interventions to the user's condition, and monitor mental health developments in real time. This has the potential to assist mental health practitioners in making faster and more informed decisions. However, the adoption of AI among practitioners remains relatively low. Influencing factors include limited technological understanding, limited infrastructure, and debates over ethical aspects and data privacy. Therefore, while AI has significant potential to improve the quality of mental health services, regulations, ethical guidelines, and synergy between technology and healthcare professionals are needed to ensure safe and effective implementation.

This study compares the performance of two deep learning models, namely Convolutional Long Short-Term Memory (ConvLSTM) and Long-term Recurrent Convolutional Network (LRCN), in the task of recognizing human activity from videos. Human activity recognition is an important field in computer vision with many applications, such as security monitoring, human-computer interaction, and social media-based video analysis. ConvLSTM is a model that combines convolution operations with long-term memory LSTM, thus capable of capturing spatial and temporal information simultaneously. This approach is ideal for processing video data sequences that have spatial and temporal dimensions. On the other hand, LRCN combines the power of spatial feature extraction from Convolutional Neural Network (CNN) and temporal sequence modeling through Recurrent Neural Network (RNN), specifically LSTM, to understand movement patterns in videos. The study used the UCF50 dataset consisting of 50 activity classes, but was limited to five classes for the focus of the experiment. The dataset was divided into 80% for training and 20% for testing, and the model was drilled for 50 epochs using early stopping to prevent overfitting. The results show that both models have high training performance. ConvLSTM achieved a training accuracy of around 98% and a validation accuracy of 90%, while LRCN achieved a training accuracy of 99.5% and a validation accuracy of 88%. Although ConvLSTM demonstrated good stability on the validation data, further testing using TikTok videos as real-world data showed that LRCN had a higher confidence level in recognizing activities, with most predictions achieving confidence scores above 80%. This difference in performance indicates that while ConvLSTM excels in generalizing on training data, LRCN is more robust to real-world data variations.

The free lunch program is a goverment initiative aimed at addressing the issue of stunting in Indonesia. This program focuses on toddlers, school-age children and pregnant women. Various opinions have emerged from the public regarding this initiative, especially through sosial media platform X (Twitter) and news portals. In this research, sentiment analysis was conducted to understand public responses to the program, whether they are positive, neutral or negative. To evaluate the accuracy of the sentiment analysis perfomed, a deep learning approach was applied using the Long Short-Term Memory (LSTM) algorithm. The results show that public sentiment varies responses, on social media X tend to be negative, while those on news portals tend to be positive toward the free lunch program in Indonesia. Through LSTM-based testing, sentiment analysis on tweet data achieved an accuracy of 88.6%, with a precision of 84.6%, recall of 88.6% and an F1-Score of 86.3%. Meanwhile, sentiment analysis on news portal data reached an accuracy of 89%, with a precision of 81.7%, recall of 89% and an F1-Score of 85.1%.

Phishing is an evolving form of cybercrime that targets users' sensitive information through URL manipulation. Conventional detection methods such as blacklists and signature-based approaches have become increasingly inadequate in addressing the dynamic variations of modern phishing attacks. This study evaluates the effectiveness of Recurrent Neural Network (RNN) variants, such Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and Gated Recurrent Unit (GRU), in detecting phishing threats based on URL data. The methodology involves a Systematic Literature Review (SLR) of scholarly publications from the past ten years, complemented by experimental implementation of the models using a public dataset from Kaggle. Literature findings show that Bi-LSTM consistently achieves the highest accuracy, up to 99%, while GRU stands out for its computational efficiency. Experimental results support these findings, with Bi-LSTM achieving an accuracy of 96.22%, GRU 96.29%, and LSTM 95.43%. Classification metrics indicate that RNN-based models perform very well in detecting benign and defacement URLs, although their performance on phishing URLs remains challenged, particularly in terms of recall. These results confirm that RNNs remain a promising approach for phishing detection systems, especially when integrated into hybrid models with complementary architectures. This study is expected to provide a foundation for developing precise and adaptive AI systems to combat increasingly sophisticated phishing threats.

Breast cancer is the most prevalent cancer among women worldwide, requiring early and accurate diagnosis to reduce mortality. This study proposes a hybrid classification pipeline that integrates Hybrid Statistical Feature Selection (HSFS) with unsupervised LSTM-guided feature extraction for breast cancer detection using the Wisconsin Diagnostic Breast Cancer (WDBC) dataset. Initially, 20 features were selected using HSFS based on Mutual Information, Chi-square, and Pearson Correlation. To address class imbalance, the training set was balanced using the Synthetic Minority Over-sampling Technique (SMOTE). Subsequently, an LSTM encoder extracted non-linear latent features from the selected features. A fusion strategy was applied by concatenating the statistical and latent features, followed by re-selection of the top 30 features. The final classification was performed using a Support Vector Machine (SVM) with RBF kernel and evaluated using 5-fold cross-validation and a held-out test set. Experimental results showed that the proposed method achieved an average training accuracy of 98.13%, F1-score of 98.13%, and AUC-ROC of 99.55%. On the held-out test set, the model reached an accuracy of 99.30%, precision of 100%, and F1-score of 99.05%, with an AUC-ROC of 0.9973. The proposed pipeline demonstrates improved generalization and interpretability compared to existing methods such as LightGBM-PSO, DHH-GRU, and ensemble deep networks. These results highlight the effectiveness of combining statistical selection and LSTM-based latent feature encoding in a balanced classification framework.

In today’s global economy, accurately predicting foreign exchange rates or estimating their trends correctly is crucial for informed investment decisions. Despite the success of standalone models like ARIMA and deep learning models like LSTM, challenges persist in capturing both linear and nonlinear dynamics in highly volatile exchange rate environments. Motivated by the limitations of these individual models and the need for more robust forecasting tools, this study proposes a hybrid ARIMA-LSTM model that integrates ARIMA’s strength in modeling linear trends with LSTM’s capability to capture nonlinear dependencies, using historical USD/NGN exchange rate data from the Central Bank of Nigeria (CBN) spanning 2001 to 2024. The research hypothesis posits that the hybrid ARIMA-LSTM model will significantly outperform standalone models in forecasting accuracy. By comparing these models against state-of-the-art approaches, the study highlights the advantages of hybridizing statistical and deep learning methods. The findings demonstrate that the hybrid model achieved the lowest Root Mean Squared Error (RMSE) of 2.216 and the highest R² of 0.998, indicating superior forecasting performance. This study fills a critical research gap by demonstrating the effectiveness of hybrid deep learning in financial time series forecasting, providing valuable insights for investors, policymakers, and financial analysts. Future research will extend this work by incorporating the latest dataset and evaluating model robustness during the recent surge in the Naira/Dollar exchange rate from 2023 to 2024.

Accurate sales forecasting is essential for stakeholders to make strategic decisions. This study aims to compare the performance of two deep learning models, namely Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN), in forecasting domestic motorcycle sales produced by AISI member manufacturers. The forecast is based on historical data from January 2021 to December 2024. The model was trained using time series data and the forecasting results for the period January to March 2025 were evaluated using the metrics Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). The results show that the LSTM model produces lower MAE and MAPE values than CNN, which shows its superiority in providing more accurate and consistent predictions. On the other hand, the CNN model has lower RMSE and MSE values, thus being able to reduce large prediction errors. By comparing the results of LSTM, CNN, and actual data forecasting, the LSTM model is more suitable for forecasting motorcycle sales in Indonesia

This research uses a deep learning-based sentiment analysis approach with several main stages, namely data collection, preprocessing, model preparation, and model building. In addition, this research also evaluates the impact of data splitting techniques on the model's performance during the training process.The evaluation results show that the LSTM-GRU model achieved the best performance on the character aspect, with an F1-score of 0.72 in the 90:10 splitting scenario. Meanwhile, the lowest F1-score was found in the editing aspect, with a value of 0.51 in the 80:20 splitting scenario. These findings indicate that the model is more effective in recognizing sentiment in narrative aspects compared to technical aspects. Furthermore, the data splitting technique significantly influences model performance, both in determining accuracy levels and in optimizing the model's effectiveness in identifying sentiment patterns more accurately.

The proliferation of Internet of Things (IoT) devices has introduced significant security challenges, necessitating robust attack detection mechanisms. This study presents a comprehensive comparative analysis of ten supervised learning algorithms for IoT attack detection and classification, addressing the critical challenge of balancing detection accuracy with practical deployment constraints. Using the CICIoT2023 dataset, encompassing data from 105 IoT devices and 33 attack types, we evaluate Naive Bayes, Artificial Neural Networks (ANN), Logistic Regression (LR), k-NN, XGBoost, Random Forest (RF), LightGBM, GRU, LSTM, and CNN algorithms based on some performance metrics. The comparative test results show superior performance to the traditional ensemble approach, with RF achieving 99.29% accuracy and leading precision (82.30%), followed closely by XGBoost with 99.26% accuracy and 79.60% precision. Deep learning approaches also demonstrate strong capabilities, with CNN achieving 98.33% accuracy and 71.18% precision, though these metrics indicate ongoing challenges with class imbalance. The analysis of confusion matrices reveals varying success across different attack types, with some algorithms showing perfect detection rates for certain attacks while struggling with others. The study highlights a crucial distinction in IoT security: while high precision remains important, the potentially catastrophic impact of missed attacks necessitates equal attention to recall metrics, as evidenced by the varying recall rates across algorithms (RF: 72.19%, XGBoost: 71.69%, CNN: 64.72%). These findings provide vital insights for developing balanced, context-aware intrusion detection systems for IoT environments, emphasizing the need to consider performance metrics and practical deployment constraints.

Social media has become a crucial part of modern life around the globe, providing users with various conveniences. However, its widespread use has also brought about new challenges, one of which is cyberbullying. This harmful issue can have serious emotional and physical effects on those targeted. Cyberbullying occurs in many areas, including sports, and soccer—a sport loved by millions—is no exception. Soccer players often face severe criticism, hate speech, and harassment on social media platforms. To tackle this problem, this study aims to create a strong model for detecting cyberbullying on the social media platform “X” using the Long Short-Term Memory (LSTM) method. By utilizing advanced machine learning techniques, the proposed model intends to identify and reduce instances of cyberbullying, helping to create a safer online space for athletes and the wider community.

The Sundanese language, once spoken by 48 million individuals, has experienced a significant decline in speakers, losing 2 million in the past decade. This decline is attributed to weakened intergenerational transmission and the dominance of more widely used languages. The challenges in developing Natural Language Processing (NLP) tools for Sundanese stem from the lack of annotated corpora, trained language models, and adequate processing tools, complicating efforts to preserve and enhance the language's usability. This research aims to address these challenges by implementing emotion classification in Sundanese text using Long Short-Term Memory (LSTM) and Bidirectional Encoder Representations from Transformers (BERT) models. The study utilizes a dataset of annotated Sundanese tweets, applying preprocessing techniques such as cleansing, stopword removal, stemming, and tokenization to prepare the data for analysis. The results indicate that the BERT model significantly outperforms the LSTM model, achieving an accuracy of approximately 80% compared to LSTM's 70%. These findings highlight the potential of advanced NLP techniques in enhancing the understanding of emotional nuances in Sundanese communication and contribute to the revitalization of the language in the digital age.