International Journal of Advances in Intelligent Informatics
Not a member yet
235 research outputs found
Sort by
Semantic-BERT and semantic-FastText model for education question classification
Full text linkQuestion classification (QC) is critical in an educational question-answering (QA) system. However, most existing models suffer from limited semantic accuracy, particularly when dealing with complex or ambiguous education queries. The problem lies in their reliance on surface-level features, such as keyword matching, which hampers their ability to capture deeper syntactic and semantic relationship in question. This results in misclassification and generic responses that fail to address the specific intent of prospective students. This study addresses, this gap by integrating semantic dependency parsing into Semantic-BERT (S-BERT) and Semantic-FastText (S-FastText) to enhance question classification performance. Semantic dependency parsing is applied to structure the semantics of interrogative sentences before classification processing by BERT and FastText. A dataset of 2,173 educational questions covering five question classes (5W1H) is used for training and validation. The model evaluation uses a confusion matrix and K-Fold cross-validation, ensuring robust performance assessment. Experimental results show that both models achieve 100% accuracy, precision, and recall in classifying question sentences, demonstrating their effectiveness in educational question classification. These findings contribute to the development of intelligent educational assistants, paving the way for more efficient and accurate automated question-answering systems in academic environments
A genetic algorithm approach to green vehicle routing: Optimizing vehicle allocation and route planning for perishable products
No full textThis paper introduces a novel approach to the Green Vehicle Routing Problem (GVRP) by integrating multiple trips, heterogeneous vehicles, and time windows, specifically applied to the distribution of bakery products. The primary objective of the proposed model is to optimize route planning and vehicle allocation, aiming to minimize transportation costs and carbon emissions while maximizing product quality upon delivery to retailers. Utilizing a Genetic Algorithm (GA), the model demonstrates its effectiveness in achieving near-optimal solutions that balance economic, environmental, and quality-focused goals. Empirical results reveal a total transportation cost of Rp. 856,458.12, carbon emissions of 365.43 kgCO2e, and an impressive average product quality of 99.90% across all vehicle trips. These findings underscore the capability of the model to efficiently navigate the complexities of real-world logistics while maintaining high standards of product delivery. The proposed GVRP model serves as a valuable tool for industries seeking sustainable and cost-effective distribution strategies, with implications for broader advancements in supply chain management
Student Major Subject Prediction Model for Real-Application Using Neural Network
Full text linkThe university admission test is an arena for students in Bangladesh. Millions of students have passed the higher secondary school every year, and only limited government medical, engineering, and public universities are available to pursue their further study. It is challenging for a student to prepare all these three categories simultaneously within a short period in such a competitive environment. Selecting the correct category according to the student's capability became important rather than following the trend. This study developed a preliminary system to predict a suitable admission test category by evaluating students' early academic performance through data collecting, data preprocessing, data modelling, model selection, and finally, integrating the trained model into the real system. Eventually, the Neural Network was selected with the maximum 97.13% prediction accuracy through a systematic process of comparing with three other machine learning models using the RapidMiner data modeling tool. Finally, the trained Neural Network model has been implemented by the Python programming language for opinionating the possible option to focus as a major for admission test candidates in Bangladesh
Chemometric classification and authentication of four Aquilaria species from essential oil profiles using GC-MS/GC-FID and ANN
Full text linkAgarwood derived from Aquilaria species is among the most valuable aromatic resources, yet frequent species misidentification hampers conservation efforts and compliance with trade regulations. This study applied a chemometric ANN framework to classify four Aquilaria species (A. malaccensis, A. beccariana, A. subintegra, and A. crassna) based on essential oil composition. A total of 720 samples (180 per species, each analyzed in triplicate) were extracted via hydro-distillation and profiled using GC–MS coupled with GC–FID. Six compounds were consistently detected, and three (δ-guaiene, 10-epi-γ-eudesmol, γ-eudesmol) were retained for classification based on ≥95% detection frequency and >0.2% relative abundance. Pearson correlation guided feature selection, and ANN models were trained using both a 70:15:15 train–validation–test split and stratified 5-fold cross-validation with 1000 bootstrap resamples. As shown in Tables 5 and 6, the optimized network achieved near-perfect performance with mean accuracy of ~99.8% (95% CI: 99.6–100.0) and precision, recall, and F1-scores all exceeding 99.5%, while bootstrapped confidence intervals were tightly bounded at 100%, confirming robustness against data leakage. These findings demonstrate that correlation-guided feature selection combined with ANN modeling enables reproducible and highly accurate species authentication, offering a practical framework for integration into agarwood quality control, conservation monitoring, and international trade compliance
Detection of errors in the Indonesian standard mushaf based on pixels to support accelerated verification
Full text linkOne effort to maintain the validity of the Al-Qur’an manuscript is the conduct of pentashihan, the analysis and verification of manuscripts by experts (pentashih). Currently, manuscript verification without translation takes 30 working days. Therefore, to support pentashih in checking the manuscript, technology is needed to speed up the pentashih process and avoid analysis errors caused by pentashih fatigue. This study conducts a writing analysis of the target manuscript by referring to the template manuscript, implementing image pre-processing stages, analysing using SSIM, and using the pixel-matching method. This method analyses and checks the manuscript's writing by comparing two block images using pixel-value analysis. Block images are the result of pre-processing the manuscript images before image matching analysis is performed. Image preprocessing consists of cropping the outer frame, cropping the inner frame, segmenting the page into row images, adjusting the margins, aligning the sizes, segmenting the row into block images, and aligning the positions between the two block images. The calculation of pixel value differences is performed at the same positions in each column and row of the template and target block images. Block image positions with pixel values ≥ 200 occur in 5 consecutive columns, adjacent rows with a distance = 1, and an SSIM value ≥ 0.9, both images meet the mismatch criteria. These findings indicate that the proposed approach provides an efficient and accurate solution for automating the verification of the Indonesian Standard Mushaf
Community preserving sparsification based on K-core for enhanced community detection in attributed networks
Full text linkCommunity detection is an important aspect of complex network analysis, especially in attribute networks where topological structure and attribute information both play a role in community formation. Traditional structure-based methods tend to result in topologically dense but semantically inconsistent communities, while attribute-based approaches can improve semantic coherence but face scalability constraints and high computational costs. On the other hand, graph sparsification techniques have been used to reduce the size of the network, but most focus on structural aspects alone and rarely consider attributes, so the quality of the resulting community is often degraded. This study proposes CPSK (Community Preserving Sparsification based on K-core), a sparsification framework that combines k-core decomposition with attribute-based side weighting. This approach is designed specifically for attribute networks, with the aim of maintaining a balance between structural reduction and community semantic consistency, while improving the efficiency of the detection process. Evaluation of the six datasets showed that CPSK consistently generates more stable and meaningful communities than existing attribute-based community detection methods, while maintaining an edge in computing efficiency on large-scale and heterogeneous networks
Predictive optimization in automotive supply chains: a BiLSTM-Attention and reinforcement learning approach
Full text linkEffective supply chain management is pivotal for enhancing customer satisfaction and driving competitiveness and profitability in the automotive service and spare parts distribution sector. Our research introduces an innovative approach, integrating game theory, BiLSTM-Attention deep learning, and Reinforcement Learning (RL) to refine supply and pricing strategies within this domain. Focusing on Moroccan automobile companies, we utilized Enterprise Resource Planning (ERP) system data to forecast customer behavior using a BiLSTM model enhanced with an Attention mechanism. This predictive model achieved a Mean Squared Error (MSE) of 0.0525 and an R² value of 0.896, indicating high accuracy and an ability to explain substantial variance in customer behavior. To further our analysis, we incorporated reinforcement learning, evaluating three algorithms: Q-learning, Deep Q-Networks (DQN), and SARSA. Our findings demonstrate SARSA's superior performance in our context, attributed to its adeptness at navigating the dynamic environment of the automotive supply chain. By synergizing the predictive power of the BiLSTM-Attention model with the strategic optimization capabilities of reinforcement learning, particularly SARSA, our study offers a comprehensive framework for automotive companies to enhance their supply chain strategies, balancing profitability and customer satisfaction effectively in a rapidly evolving industry secto
Academic expert finding using BERT pre-trained language model
Full text linkAcademic expert finding has numerous advantages, such as: finding paper-reviewers, research collaboration, enhancing knowledge transfer, etc. Especially, for research collaboration, researchers tend to seek collaborators who share similar backgrounds or with the same native languages. Despite its importance, academic expert findings remain relatively unexplored within the context of Indonesian language. Recent studies have primarily relied on static word embedding techniques such as Word2Vec to match documents with relevant expertise areas. However, Word2Vec is unable to capture the varying meanings of words in different contexts. To address this research gap, this study employs Bidirectional Encoder Representations from Transformers (BERT), a state-of-the-art contextual embedding model. This paper aims to examine the effectiveness of BERT on the task of academic expert finding. The proposed model in this research consists of three variations of BERT, namely IndoBERT (Indonesian BERT), mBERT (Multilingual BERT), and SciBERT (Scientific BERT), which will be compared to a static embedding model using Word2Vec. Two approaches were employed to rank experts using the BERT variations: feature-based and fine-tuning. We found that the IndoBERT model outperforms the baseline by 6–9% when utilizing the feature-based approach and shows an improvement of 10–18% with the fine-tuning approach. Our results proved that the fine-tuning approach performs better than the feature-based approach, with an improvement of 1–5%. It concludes by using IndoBERT, this research has shown an improved effectiveness in the academic expert finding within the context of Indonesian language
CMT-CNN: colposcopic multimodal temporal hybrid deep learning model to detect cervical intraepithelial neoplasia
Full text linkCervical cancer poses a significant threat to women's health in developing countries, necessitating effective early detection methods. In this study, we introduce the Colposcopic Multimodal Temporal Convolution Neural Network (CMT-CNN), a novel model designed for classifying cervical intraepithelial neoplasia by leveraging sequential colposcope images and integrating extracted features with clinical data. Our approach incorporates Mask R-CNN for precise cervix region segmentation and deploys the EfficientNet B7 architecture to extract features from saline, iodine, and acetic acid images. The fusion of clinical data at the decision level, coupled with Atrous Spatial Pyramid Pooling-based classification, yields remarkable results: an accuracy of 92.31%, precision of 90.19%, recall of 89.63%, and an F-1 score of 90.72. This achievement not only establishes the superiority of the CMT-CNN model over baselines but also paves the way for future research endeavours aiming to harness heterogeneous data types in the development of deep learning models for cervical cancer screening. The implications of this work are profound, offering a potent tool for early cervical cancer detection that combines multimodal data and clinical insights, potentially saving countless lives
Computer-aided pulmonary disease diagnosis using lung ultrasound video
Full text linkThe development of a machine learning-based computer-aided diagnosis (CAD) system implemented for processing lung ultrasound images will greatly assist doctors in making decisions in diagnosing lung diseases. The learning method of the classifier model used in the computer-aided diagnosis system will affect the system's accuracy in diagnosing lung disease. Determining variables in the classifier and image pre-processing stages requires special attention to obtain a highly accurate classifier model. This study presents the development of a machine learning-based CAD as an add-on tool to classify lung disease based on a lung ultrasound (LUS) video. The main steps in this study are capturing the LUS videos and converting them into images, image pre-processing for speckle noise removal, image contrast and brightness enhancement, feature extraction, and the classification stage. In this study, three learning algorithm models, namely Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Naïve Bayes (NB), were used to classify images into three categories, namely healthy conditions, pneumonia, and COVID-19. The performance of the three classifier models is compared to each other to obtain the best classifier model. The experimental results demonstrate the superiority of the suggested strategy utilizing the SVM classifier. Based on experimental data using 2,149 lung images for three classes and 20 texture feature sets, the SVM has an accuracy of 98.1%, the KNN is 94.7%, and the Gaussian NB is 79.6%. The model with the highest accuracy will be used to develop the computer-aided diagnosis (CAD) system