Jurnal Teknik Informatika dan Sistem Informasi

A Comparative Study of Box–Jenkins and Machine Learning Model for Water Level Prediction

Thu, 27 Aug 2026 00:00:00 +0000

The Pasar Ikan Water Gate, located in North Jakarta, is a strategic area that is prone to coastal flooding (rob). The water level in this area is influenced by a combination of tides, high rainfall, and poor drainage systems, making it a frequent early indicator of potential flooding. This study uses daily water level measurement data from the Fish Market Water Gate from September 1 to October 24, 2024. The data patterns show that the data is quite fluctuating and exhibits seasonal patterns. Therefore, this study aims to compare the performance of the classical method, seasonal autoregressive integrated moving average (SARIMA), with the machine learning-based method, long short-term memory (LSTM). The analysis results show that the best SARIMA model is the ARIMA(1,0,3)×(1,2,1)_24 model with a mean absolute percentage error (MAPE) of 10.88% and a root mean square error (RMSE) of 25.45. However, this model has weaknesses due to unmet model assumptions, namely the assumptions of normality and homogeneity of residual variance. The LSTM modeling results indicate that the model with a combination of 50 epochs and a batch size of 1 produces the best LSTM model with MAPE and RMSE of 2.58% and 5.60, respectively. The LSTM method's prediction results are better than the SARIMA method's because they still follow the actual data pattern and show an upward trend, whereas the SARIMA method's forecasting results fail to follow this pattern and do not meet the assumptions.

Transformation of Information Services Using Retrieval-Augmented Generation and Large Language Model

Muhamad Komarudin, Chelly Sabrina, Yessy Mulyani, Puput Budi Wintoro — Thu, 27 Aug 2026 00:00:00 +0000

The openness of access to public information is regulated in Law No. 14 of 2008 on Public Information Disclosure as a key element in realizing transparency and good governance. At the University of Lampung, Pejabat Pengelola Informasi dan Dokumentasi (PPID) provides various public data through its official website. However, users often face difficulties in finding specific information due to the dispersed nature of documents and the limitations of search features. To address this issue, a Telegram chatbot powered by Artificial Intelligence was developed using the Large Language Model (LLM) Qwen2.5 VL 72B Instruct integrated with the Retrieval-Augmented Generation (RAG) architecture. Data were collected from the University of
Lampung’s PPID website and organized into a structured dataset of 375 curated question–answer pairs derived from 47 public information entries. These were processed through tokenization, embedding, and indexing using the FAISS vector database to support semantic search. The evaluation was carried out using RAGAS metrics (faithfulness, answer relevance, and context recall)
with a minimum threshold of 0.8, along with a performance comparison against a pure LLM. The results show that the developed system successfully surpassed the RAGAS minimum threshold with a remarkably low average response latency of 0.69 seconds and achieved higher accuracy compared to the pure LLM, with a significant reduction in answer hallucinations. Furthermore, usability testing using the Chatbot Usability Questionnaire (CUQ) produced an average score of 90.62, indicating an excellent user experience. This study concludes that integrating LLM and RAG in a chatbot effectively improves accuracy, relevance, and the ease of real-time access to public information.

Optimizing Banking Stock Prediction with Single Layer Long Short-Term Memory

Beny Beny, Herti Yani, Niansyah Eko Putra — Thu, 27 Aug 2026 00:00:00 +0000

The inherent volatility of the stock market and increasing investor engagement, especially in the banking sector, highlight the critical need for accurate stock price prediction. This research develops and evaluates a single-layer Long Short-Term Memory (LSTM) model to forecast the closing prices of six major Indonesian banking stocks: PT Bank Central Asia Tbk (BBCA.JK), PT Bank Rakyat Indonesia Tbk (BBRI.JK), PT Bank Mandiri Tbk (BMRI.JK), PT Bank Negara Indonesia Tbk (BBNI.JK), PT Bank MEGA Tbk (MEGA.JK), and PT Bank Tabungan Negara Tbk (BBTN.JK). A single-layer LSTM architecture was systematically tested using 4 distinct hyperparameter including number of neurons, batch size, optimizers, and learning rates. Model performance was assessed using Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). Experimental results demonstrate that a single-layer LSTM model with 100 neurons achieved superior accuracy with lowest MAPE of 1.95% and RMSE of 106.54 on BBNI stock compared to configurations with 50 or 200 neurons. Furthermore, with 100 neuron and 0.001 learning rate setups, the Adam and Adamax optimizers consistently outperformed RMSProp across all stocks prediction. These findings provide practical guidance on hyperparameter tuning for LSTM-based stock prediction in the Indonesian banking industry.

Clothing Line Strategy Based on Business Model Canvas and Customer Relationship Management

Thu, 27 Aug 2026 00:00:00 +0000

The pace of technological development and technology adoption in the business sector is a challenge for business people, so it is necessary to adjust conventional business models towards digital. The research object of Destiny Project Salatiga which is engaged in the Clothing Line business uses a conventional sales model in carrying out operations. This is still considered ineffective to reach sales marketing, so a strategy and implementation of technology use is needed. The purpose of the study is to analyze business strategies using the Business Model Canvas and the implementation of business processes with the concept of E-CRM in order to recommend digital-based business processes. The research used a descriptive qualitative method by collecting data through interviews. The results of conventional business model research then business strategy analysis in the form of business process visualization based on the Business Model Canvas framework and knowing the advantages, weaknesses, threats and business opportunities, then the E-CRM-based implementation stage produces a digital business process design and display design in mobile form. Thus providing quality and comfortable service to customers, reaching a wide market, facilitating feedback between customers and business people, it is hoped that later good relations will continue to be established. The adoption of technology in the Destiny Project Salatiga business, which was initially conventional and then transformed into digital, is able to support and support business operations. Research contributes to the field of entrepreneurship and technology to support digital business transformation.

Visibility Prediction Using Long Short-Term Memory on Multivariate Spatio-Temporal Data

Thu, 27 Aug 2026 00:00:00 +0000

Visibility prediction at airports is a crucial factor in maintaining flight safety and operational efficiency. This study aims to develop a visibility prediction model at Juanda International Airport by utilizing multivariate Automatic Weather Observation System (AWOS) data based on the Long Short-Term Memory (LSTM) approach. The dataset consists of one-minute weather observations throughout 2022 from three observation points, namely point 10, point 10–28, and point 28. The research process includes data preprocessing, missing value imputation using continuous and vector interpolation, dimensionality reduction with Principal Component Analysis (PCA), sequence formation of 168 time steps, and training of a two-layer LSTM model with dropout regularization. The evaluation results show that the model is able to provide accurate and stable predictions, with a Mean Absolute Error (MAE) of 0.0617 on validation data and 0.0618 on test data. Meanwhile, the Mean Squared Error (MSE) obtained is 0.0170 on validation data and 0.0150 on test data, while the Root Mean Square Error (RMSE) is 0.1305 on validation data and 0.1225 on test data. The small differences between validation and test results confirm the good generalization capability of the model without significant overfitting, while the use of PCA has proven effective in enhancing performance by highlighting the most relevant features. Overall, this study demonstrates that the multivariate LSTM approach using AWOS data is effective for supporting automatic visibility prediction systems at airports and opens opportunities for the development of data-driven early warning systems.

Parameters Evaluation of Flower Pollination Neural Network for Malaria and Dengue Detection

Thu, 27 Aug 2026 00:00:00 +0000

As the capital city of East Nusa Tenggara Province (NTT), Kupang City faces serious issues related to waste management. The accumulation of unmanaged waste leads to water stagnation and creates ideal breeding grounds for mosquitoes, thereby increasing the risk of disease transmission such as Dengue Fever (DBD) and Malaria. Individuals aged 0–20 years and females tend to have higher rates of DBD, while malaria most commonly affects young children who have not yet developed immunity and pregnant women whose immunity is reduced due to pregnancy. To address this challenge, this study proposes a novel approach: the Flower Pollination Neural Network (FPNN) for classifying DBD and malaria cases. FPNN is a hybrid model combining Neural Networks (NN) optimized using the Flower Pollination Algorithm (FPA). FPA is a metaheuristic algorithm that often employs stochastic search. It is capable of optimizing the weights of NN, avoiding local optima traps, and minimizing errors in NN. This study uses 10-fold cross-validation to evaluate FPNN’s ability to generalize to unseen data. The best results were achieved using a parameter combination of switch probability 0.7, 15 neurons in the hidden layer, and a population size of 25, yielding an accuracy of 86.4% and an F1-score of 72.312%. Although this parameter combination does not outperform all others in terms of accuracy, it consistently delivers good accuracy and demonstrates strong data generalization capabilities. Furthermore, it achieved the highest F1-score, indicating a good balance between pattern recognition during training and prediction performance on unseen data.

Enhancing DeepMelaNet Accuracy Through Hairline Removal and Image Contrast Enhancement

Nurvianto Akbar Ikhsani, Yufis Azhar — Thu, 27 Aug 2026 00:00:00 +0000

Melanoma is one of the most dangerous types of skin cancer, originating from melanocytes that produce melanin as skin pigment. This disease can develop aggressively and has a high risk of spreading to other organs if not detected early. Accurate diagnosis makes it possible to prevent this disease and avoid death from melanoma. This study aims to optimize image-based melanoma classification results by utilizing the DeepMelaNet model through the integration of two main preprocessing techniques, namely hair line removal and contrast enhancement using the CLAHE method. The problem of inconsistent dermoscopy image quality is an obstacle in identifying relevant clinical details, so preprocessing techniques are needed to improve the generalization and reliability of the model. This study also applies several augmentations which are then further processed with DeepMelaNet-based classification experiments. Evaluations are carried out on various preprocessing combination scenarios to assess the impact on classification accuracy. The main findings show that the sequence of applying contrast enhancement followed by hairline removal and optimal learning rate adjustment can improve validation accuracy to 96.5 percent, surpassing the standalone DeepMelaNet approach and other combinations while producing the most stable training performance. This study confirms that the selection and adjustment of preprocessing strategies play a crucial role in improving the accuracy and generalization of deep learning models for melanoma classification tasks, so that this method can be recommended as a standard in Deep Learning-based dermoscopy medical image analysis.

Reconstruction of Traditional Dance Movements with Hidden Markov Model Multimodal

Anang Kukuh Adisusilo, Emmy Wahyuningtyas, Teguh Pribadi Ikhsan — Thu, 27 Aug 2026 00:00:00 +0000

The preservation of traditional dance in the digital era requires a computational representation that not only captures bodily motion, but also maintains the stylistic continuity and cultural embodiment embedded within its choreographic structure. This study presents a comparative evaluation of three emission models in Hidden Markov Models (HMM)—Single-Gaussian, Gaussian Mixture Model (GMM-HMM), and Multinomial HMM—for reconstructing the Bedoyo Majapahit classical dance using markerless motion capture data. The recorded 3D skeleton sequence consists of 3,341 frames and 33 joint coordinates per frame, which were normalized and filtered using a low-pass smoothing technique. Principal Component Analysis (PCA) was then applied to reduce dimensionality while preserving the primary spatial-temporal variance of the motion features. The reconstruction performance was assessed using three quantitative metrics: Mean Squared Error (MSE) to measure geometric fidelity, Dynamic Time Warping (DTW) to evaluate temporal consistency, and Fréchet distance to assess global trajectory similarity. The results show that the GMM-HMM significantly outperforms the other two models across all evaluation metrics, demonstrating superior stability, curvature preservation, and alignment with the original motion pattern. The findings highlight that traditional choreography exhibits inherently multimodal motion characteristics, which are better captured by Gaussian mixture emissions rather than unimodal or discretized representations. Beyond numeric superiority, the GMM-HMM also preserves stylistic curvature that reflects key embodied principles of Javanese dance—wiraga (embodied form), wirama (temporal flow), and wirasa (expressive intention). This study provides a methodological foundation for culturally faithful digital preservation and serves as a basis for future AI-assisted interactive learning systems for traditional dance heritage.

Transformer Model for Long-Term Stock Market Volatility Prediction

Thu, 27 Aug 2026 00:00:00 +0000

Predicting long-term stock market volatility in emerging stock markets is a significant challenge due to the data’s inherent non-linear and complex nature. This study aims to design and evaluate a Transformer-based deep learning architecture to enhance the accuracy of long-term volatility prediction, with a case study on the Indonesia Stock Exchange Composite Index (IDX Composite). The proposed method integrates multi-modal time-series data, including historical market data (IDX Composite, blue-chip stocks), key macroeconomic indicators (interest rates, inflation), and cross-aset data (USD/IDR). The Transformer architecture, leveraging its self-attention mechanism, is employed to capture complex long-range temporal dependencies within the data. The model was trained to predict future realized volatility over a 21-day horizon. Its performance was rigorously evaluated against popular baseline architectures, namely Long Short-Term Memory (LSTM) and GARCH(1,1), using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) metrics. The evaluation results on the test set demonstrate that the proposed Transformer model consistently outperforms both LSTM and GARCH models, indicated by lower RMSE and MAE values. This finding highlights the superior capability of the Transformer architecture in modeling complex non-linear relationships and long-term dependencies from multi-modal financial data. In conclusion, the Transformer architecture proves to be a robust and effective framework for long-term volatility prediction, offering a promising solution for risk analysis and strategic investment decision-making in emerging markets.

Analysis of Instagram Content Optimization for a Small Toasted Bread Business

Adelia Adelia, Diana Trivena Yulianti, Doro Edi, Olivia Justine Audia — Thu, 27 Aug 2026 00:00:00 +0000

Technological advances and easy access to technology in today's society can be utilized by small businesses, one of which is a toast business. This technology is utilized by using Instagram as a social media marketing platform. Toast businesses can utilize Instagram to promote their products by creating content that can attract audience attention. Content optimization efforts are carried out to create content that can attract audience attention, which can be determined by calculating the engagement rate. The level of audience engagement on each piece of content in a toast business can be increased by several factors, namely paying attention to the quality and type of content, posting time, and posting frequency. Analysis is carried out on each piece of content created, paying attention to aspects of increasing the engagement rate of each piece of content and can expand the market reach for toast businesses.

Multi-Intent Chatbot for New Student Admission Information Using Artificial Neural Networks

M Thoriq Panca Mukti, Safrin Safrin, Moh Sa'id Ramdhoni — Thu, 27 Aug 2026 00:00:00 +0000

Digital transformation in higher education services requires the provision of New Student Admission information (PMB) services that are fast, accurate, and available on an ongoing basis. However, conventional PMB services still face the problem of limited service time, staff workload, and difficulty in handling user inquiries that vary and are multi-intent. This study aims to implement PMB information service chatbot based on Artificial Intelligence that is able to handle multi-intent questions effectively. The proposed approach integrates IndoBERT Pre-trained language model as semantic feature extractor with Artificial Neural Network (ANN) as intent classifier. The research Dataset consisted of 1,064 question variations grouped into 21 intent categories. The test results showed that the model achieved an accuracy rate of 96%, with the value of precision, recall, and F1-score of 0.96, respectively. The implementation of the system proves that the chatbot is able to provide relevant responses in both single-intent and multi-intent scenarios. Thus, this approach is effective in improving the quality and flexibility of chatbot-based PMB information services.

Tomato Grade Classification Using You Only Look Once for Post-Harvest Sorting

Anggoro Panji Sulistyo, Magdalena Ariance Ineke Pakereng — Thu, 27 Aug 2026 00:00:00 +0000

Tomatoes are one of the horticultural commodities with high economic value; however, quality grading in small-scale agricultural environments is still predominantly performed manually, leading to inconsistency, subjectivity, and high labor dependency. This study proposes a tomato grade classification system based on YOLOv8-Large to optimize post-harvest sorting at the farmer level. The dataset consists of 683 field images annotated using a polygon tool on the Roboflow platform, divided into 576 training images, 64 validation images, and 43 testing images. The model was trained using the AdamW optimizer with a batch size of 16 for 50 epochs. Evaluation results show a mAP@0.5 of 97.5% and a mAP@0.5:0.95 of 95.3%, with precision and recall values of 93.3% and 91.6%, respectively. During testing, the system processed approximately 40 tomatoes in ±2.5 seconds per container, improving sorting efficiency by approximately 97.2% compared to the manual method, which required around 90 seconds. These results demonstrate the potential of artificial intelligence to support precision agriculture and its integration into automated sorting systems such as conveyor-based solutions.

Machine Learning Framework for Early Warning of Rating Declines in Tourism Accommodations

Thu, 27 Aug 2026 00:00:00 +0000

This study develops an early warning system framework to predict rating declines in tourism accommodations using machine learning with a validated synthetic data approach. The main challenge addressed is that rating decreases on digital platforms are typically identified only after they occur, resulting in delayed managerial responses and negative impacts on business reputation. The proposed predictive framework integrates customer review sentiment analysis, topic modeling, and temporal aggregation to detect early signals of rating decline one period in advance. Synthetic data were constructed based on the characteristics of the public Bali Hotel Reviews dataset and Google Reviews API documentation, followed by comprehensive statistical validation. Model evaluation indicates that the Random Forest algorithm achieves the best performance in predicting rating declines. The results further show that topic probability, average rating, and minimum rating are the most influential predictors. The developed framework provides interpretable and actionable early warnings through an interactive dashboard with minimal infrastructure requirements.

Tomato Leaf Disease Classification Using the K-Nearest Neighbor Method with Watershed Segmentation

Madadina Adilah Pamuji, Rinci Kembang Hapsari — Thu, 27 Aug 2026 00:00:00 +0000

Tomato leaf diseases are one of the main factors causing a decline in tomato crop productivity and agricultural yields. Manual identification of leaf diseases is subjective, time-consuming, and prone to inconsistency. Therefore, this study proposes an automated tomato leaf disease classification system based on digital image processing and machine learning by combining watershed segmentation, HSV color features, Gray Level Co-occurrence Matrix (GLCM) texture features, and the K-Nearest Neighbor (KNN) algorithm. The dataset used consists of 8,111 single-leaf tomato images covering nine disease classes and one healthy class, as well as an additional multi-leaf dataset for segmentation evaluation. The research process includes image preprocessing, leaf segmentation using the watershed method, feature extraction using HSV and GLCM, and classification using KNN with Euclidean distance. Experimental results show that the proposed system achieved an accuracy of 87.08% on training data, 82.96% on validation data, and 81.98% on test data. Furthermore, the watershed segmentation method successfully separated individual leaves in multi-leaf images, enabling effective per-leaf classification. These results indicate that the combination of handcrafted features and the KNN algorithm remains effective and computationally efficient for tomato leaf disease classification tasks.

Evaluation of Acceptance of Recommender System-Based Academic Information System Using WebQUAL Method

Thu, 27 Aug 2026 00:00:00 +0000

This study aims to evaluate the acceptance level of a web-based recommender academic information system at the Informatics Engineering Department of Halu Oleo University using the WebQual 4.0 method. The evaluation focuses on three main dimensions: usability, information quality, and interaction quality. A quantitative approach was employed by distributing questionnaires to 40 active system users. The collected data were analyzed descriptively and complemented with reliability, validity, and multicollinearity tests to ensure the consistency of the instruments. The results show a Cronbach’s Alpha value of 0.743, indicating good reliability, and no multicollinearity issues were found in the regression model. Based on the evaluation, the system obtained high average scores in usability (4.4–4.5), information quality (4.4–4.7), and interaction quality (4.4–4.6). These findings indicate that the web-based recommender system significantly improves academic service efficiency through easy access, faster data management, and better user interaction. The results serve as a foundation for developing a more responsive, secure, and interactive academic information system to enhance service quality and user satisfaction.

Implementing Fine-Tuned IndoBERT for Detecting Potential Indonesian Hoax News

Viny Christanti Mawardi, Henokh Mikhael Kristiantan — Thu, 27 Aug 2026 00:00:00 +0000

The spread of fake information or hoaxes on Indonesian digital platforms has become a serious threat to social stability and public literacy. This study aims to design and implement a web-based application capable of automatically detecting potential Indonesian hoax news. The proposed method utilizes a Natural Language Processing approach by applying fine-tuning techniques to the pre-trained IndoBERT model (IndoBERT Base P1). The model was specifically trained using an Indonesian news text dataset to classify and recognize the linguistic patterns of disinformation. Key findings demonstrate that the fine-tuned IndoBERT model was successfully integrated into the web system architecture, delivering excellent classification performance with accuracy, precision, and F1-score reaching 98.6%. Furthermore, the system is capable of processing input text and displaying the confidence score percentage in real-time to the user. In conclusion, the implementation of fine-tuned IndoBERT in a web application has proven to be an effective and reliable early verification tool. The presence of this system is expected to assist the public in filtering information and contribute to suppressing the spread of hoaxes within the Indonesian digital space.