Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi)
Not a member yet
1071 research outputs found
Sort by
Comparative Performance of ResNet Architectures for Toraja Carving Image Classification with Data Augmentation
No full textThe complexity of the motifs and large number of different patterns make the classification of Toraja carvings challenging. The objective of this study is to develop a Convolutional Neural Network automatic classification model using a comparative analysis of the performance of three ResNet architectures. Data augmentation techniques were used to enrich the diversity of the training samples and improve the robustness of the model. The experimental results showed that ResNet101V2 had the highest validation accuracy, which was greater than 97%, followed by ResNet50V2 with more than 96%, and finally, ResNet152V2 with more than 94.74%. These test results indicate that the ResNet101V2 architecture has a better classification performance for complex motifs, with a good balance between precision and recall. However, the confusion matrix and per-class performance metrics indicated that motifs with high similarity, such as Paqdon-Bolu and Paqtedong, remained challenging. This study demonstrated that deeper CNN architectures and data augmentation techniques are effective in improving the classification accuracy of complex carving patterns. Further research should explore hybrid or advanced augmentation methods to improve the overall robustness and accuracy of the model
Application of Formal Concept Analysis and Clustering Algorithms to Analyze Customer Segments
No full textBusiness development cannot be separated from relationships with customers. Understanding customer characteristics is important both for maintaining sales and even for targeting new customers with appropriate strategies. The complexity of customer data makes manual analysis of the customer segments difficult, so applying machine learning to segment the customer can be the solution. This research implements K-Means and GMM algorithms for performing clustering based on the Transaction data transformed to the Recency, Frequency, and Monetary (RFM) data model, then implements Formal Concept Analysis (FCA) as an approach to analyzing the customer segment after the class labeling. Both K-Means and GMM algorithms recommended the optimal number of clusters as the customer segment is four. The FCA implementation in this study further analyzes customer segment characteristics by constructing a concept lattice that categorizes segments using combinations of High and Low values across the RFM attributes based on the median values, which are High Recency (HR), Low Recency (LR), High Frequency (HF), Low Frequency (LF), High Monetary (HM), and Low Monetary (LM). This characteristic can determine the customer category; for example, a customer that has HM and HR can be considered a loyal customer and can be the target for a specific marketing program. Overall, this study demonstrates that using the RFM data model, combined with clustering algorithms and FCA, is a potential approach for understanding MSME customer segment behavior. However, special consideration is necessary when determining the FCA concept lattice, as it forms the foundation of the core analytical insights
Large Language Model-Based Extraction of Logic Rules from Technical Standards for Automatic Compliance Checking
No full textIn this research, we design logic rules as a representation of technical standards documents related to ship design, which will be used in automatic compliance checking. We present a novel design of logic rules based on a general pattern of technical standards’ clauses that can be produced automatically from text using a large language model (LLM). We also present a method to extract said logic rules from text. First, we design data structures to represent the technical standards and logic rules used to process the data. Second, the representation of technical standards is produced manually and tested to ensure that it can give the same conclusion as human judgment regarding compliance. Third, a variation of prompting methods, namely pipeline method and few-shot prompting, is given to LLM to instruct it to extract logic rules from text following the design. Evaluation against the logic rules produced shows that the pipeline method gives an accuracy score of 0.57, a precision of 0.49, and a recall of 0.62. On the other hand, logic rules extracted using few-shot prompting have an accuracy score of 0.33, precision of 0.43, and recall of 0.5. These results show that LLM is able to extract a logic rule representation of technical standards. Furthermore, the representation resulting from the prompting technique that utilizes the pipeline method has a better performance compared to the representation resulting from few-shot prompting
Improving Government Helpdesk Service With an AI-Powered Chatbot Built on the Rasa Framework
No full textHelpdesk services are an important component in supporting Information Technology (IT) services. The helpdesk operates as the initial interface for managing and resolving concerns. Helpdesk helps user to get solutions when facing problems while using an IT service. This research focuses on the impact of artificial intelligence (AI)-powered chatbots on the performance of the initial response of government helpdesk services. The chatbot is designed to improve service performance by quickly identifying and classifying reported issues and automatically responding to messages, enabling faster responses. The research proposed a new System Design of a helpdesk system with an AI-based chatbot. The data used comes from Telegram group chat logs, exported in JSON format. We find that the Rasa NLU model with DIET Classifier successfully achieved an accuracy rate of 0.825 in classifying intents, with the precision value of 0.838, recall of 0.829, and F1 score of 0.821 using a Rasa model with cross-validation, where folds is 5 in evaluation. And initial response time was highly improved after using chatbot artificial intelligence from more than 3 hours on the telegram group helpdesk based to an average of 2.15 seconds. These research results suggest AI-Chatbot-based ability to assist the helpdesk team in handling user queries and reports, and improving initial time response
Word2Vec Approaches in Classifying Schizophrenia Through Speech Pattern
No full textSchizophrenia is a chronic brain disorder characterized by symptoms such as delusions, hallucinations, and disorganized speech, posing significant challenges for accurate diagnosis. This research investigates an innovative Natural Language Processing (NLP) framework for classifying the speech patterns of schizophrenia patients using Word2Vec, with the aim of determining whether there are significant differences between the two features. The dataset comprises speech transcriptions from 121 schizophrenia patients and 121 non-schizophrenia participants collected through structured interviews. This study compares two Word2Vec architectures, Continuous Bag-of-Words (CBOW) and Skip-Gram (SG), to determine their effectiveness in classifying schizophrenia speech patterns. The results indicate that the SG architecture, with hyperparameter tuning, produces more detailed word representations, particularly for low-frequency words. This approach yields more accurate classification results, achieving an F1-score of 93.81%. These results emphasize the effectiveness of the framework in handling structured and abstract linguistic patterns. By utilizing the advantages of both static and contextual embedding, this approach offers significant potential for clinical applications, providing a reliable tool for improving schizophrenia diagnosis through automated speech analysis
Efficient Hybrid Network with Prompt Learning for Multi-Degradation Image Restoration
No full textImage restoration aims to repair degraded images. Traditional image restoration methods have limited generalization capabilities due to the difficulty in dealing with different types and levels of degradation. On the other hand, contemporary research has focused on multi-degradation image restoration by developing unified networks capable of handling various types of degradation. One promising approach is using prompts to provide additional information on the type of input images and the extent of degradation. Nonetheless, all-in-one image restoration requires a high computational cost, making it challenging to implement on resource-constrained devices. This research proposes a multi-degradation image restoration model based on PromptIR with lower computational cost and complexity. The proposed model is trained and tested on various datasets yet it is still practical for deraining, dehazing, and denoising tasks. By unification convolution, transformer, and dynamic prompt operations, the proposed model successfully reduces FLOPs by 32.07% and the number of parameters by 27.87%, with a comparable restoration result and an SSIM of 34.15 compared to 34.33 achieved by the original architecture for the denoising task
The Impact of Cancer on Poverty: An Analytical Study Using Big Data and OLS Regression
Full text linkCancer is one of the leading causes of death worldwide and has a significant impact on the economic condition of families, especially in developing countries. High medical costs and loss of work productivity often push families of patients with cancer into poverty. This study aimed to analyze the relationship between cancer mortality rates and poverty levels using the Ordinary Least Squares (OLS) regression method and big data covering various socio-economic indicators. The data in this study include cancer mortality rates and other socioeconomic indicators, which were then analyzed using the OLS regression method to understand the quantitative relationship between the two variables. The results of the analysis show a positive correlation between cancer mortality rates and increasing poverty, with the regression model explaining 73.8% of the variation in the target variable. The regression model demonstrated strong explanatory power and minimal error, with an R-squared value of 0.738, indicating that 73.8% of the data variability was explained by the model. Model quality was supported by low AIC (19070.4) and BIC (19110.4) values. Linearity was confirmed by a significant F-statistic of 1314.0 (p < 0.01), suggesting a robust linear relationship between independent and dependent variables. All parameters exhibited statistical significance (p < 0.05) at the 95% confidence level, with mean residuals close to zero, satisfying the unbiased expectation assumption. Although the model results show good performance, the model's estimators show low variance, as evidenced by small standard errors (e.g., Incidence_Rate: 0.009, Med_Income: 1.89e-05) and a Durbin-Watson statistic of 1.725, indicating no autocorrelation. These metrics collectively confirmed the reliability and stability of the regression model
Prediction of Financial Distress in Retail Companies Using Long-Short Term Memory (LSTM)
Full text linkFinancial distress is a condition in which an entity struggles to meet its debt and operating obligations.. Financial distress can lead to bankruptcy or company closure if corrective action is not taken. This study aims to forecast financial distress in retail companies by utilizing key financial ratios, including Total Asset Turnover (TATO), Current Ratio (CR), Return on Assets (ROA), and Debt-to-Equity Ratio (DER). The analysis is based on secondary data from Indonesian retail companies listed on the Indonesia Stock Exchange (IDX) during the 2022–2024 period. The dataset exhibited missing values and class imbalance, which were addressed using mean imputation and the Synthetic Minority Oversampling Technique (SMOTE), respectivelyTo perform predictions, a Long Short-Term Memory (LSTM) model was implemented. The integration of SMOTE contributed to enhanced detection of the minority class; however, it was accompanied by a slight reduction in overall predictive accuracy. The model demonstrated a performance accuracy of 86%, with a recall rate of 85%, a precision of 100%, and an F1-score of 92%
Health Risk Classification Using XGBoost with Bayesian Hyperparameter Optimization
Full text linkHealth risk classification is important. However, health risk classification is challenging to address using conventional analytical techniques. The XGBoost algorithm offers many advantages over the traditional methods for risk classification. Hyperparameter Optimization (HO) of XGBoost is critical for maximizing the performance of the XGBoost algorithm. The manual selection of hyperparameters requires a large amount of time and computational resources. Automatic HO is needed to avoid this problem. Several studies have shown that Bayesian Optimization (BO) works better than Grid Search (GS) or Random Search (RS). Based on these problems, this study proposes health risk classification using XGBoost with Bayesian Hyperparameters Optimization. The goal of this study is to reduce the time required to select the best XGBoost hyperparameters and improve the accuracy and generalization of XGBoost performance in health risk classification. The variables used were patient demographics and medical information, including age, blood pressure, cholesterol, and lifestyle variables. The experimental results show that the proposed approach outperforms other well-known ML techniques and the XGBoost method without HO. The average accuracy, precision, recall and f1-score produced by the proposed method are 0.926, 0.920, 0.928, and 0.923, respectively. However, improvements are needed to obtain a faster and more accurate method in the future.
 
DiG-MFV: Dual-integrated Graph for Multilingual Fact Verification
Full text linkThe proliferation of misinformation in political domains, especially across multilingual platforms, presents a major challenge to maintaining public information integrity. Existing models often fail to effectively verify claims when the evidence spans multiple languages and lacks a structured format. To address this issue, this study proposes a novel architecture called Dual-integrated Graph for Multilingual Fact Verification (DiG-MFV), which combines semantic representations from multilingual language models (i.e., mBERT, XLM-R, and LaBSE) with two graph-based components: an evidence graph and a semantic fusion graph. These components are processed through a dual-path architecture that integrates the outputs from a text encoder and a graph encoder, enabling deeper semantic alignment and cross-evidence reasoning. The PolitiFact dataset was used as the source of claims and evidence. The model was evaluated by using a data split of 70% for training, 20% for validation, and 10% for testing. The training process employed the AdamW optimizer, cross-entropy loss, and regularization techniques, including dropout and early stopping based on the F1-score. The evaluation results show that DiG-MFV with LaBSE achieved an accuracy of 85.80% and an F1-score of 85.70%, outperforming the mBERT and XLM-R variants, and proved to be more effective than the DGMFP baseline model (76.1% accuracy). The model also demonstrated stable convergence during training, indicating its robustness in cross-lingual political fact verification tasks. These findings encourage further exploration in graph-based multilingual fact verification systems