Garuda - Garba Rujukan Digital

Improving Brain Tumor Image Segmentation Accuracy Based on Residual Network (ResNet) Using Nearest Neighbor Upsampling Muhammad Afifudin; Endang Sugiharti
Recursive Journal of Informatics Vol. 4 No. 1 (2026): March 2026
Publisher : Universitas Negeri Semarang

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.15294/rji.v4i1.12010

Abstract. Brain tumors are a critical disease due to the abnormal growth of cells in the brain, which can damage surrounding normal cells and increasing the risk of death in patients. With advancements in technology, artificial intelligence can be developed to segment brain tumor areas, aiding medical professionals in identifying tumor characteristics and determining appropriate treatment plans. Convolutional Neural Network (CNN) models can be utilized for segmentation tasks because their ability to classify each pixel of an image, assign specific labels, and map them into homogeneous groups. To enhance the capability of CNNs against the possibility of vanishing gradients, the Residual Network (ResNet) architecture can be applied to the segmentation model. The use of ResNet provides additional capability for the network to choose between the training results in the current epoch or skip to the next network when the training results approach the identity value. However, ResNet also reduces the scale of images and feature maps during downsampling operations, sacrificing spatial resolution. This study proposes the implementation of the Nearest Neighbor Upsampling method on ResNet to improve the model's accuracy in the task of MRI brain tumor segmentation. Purpose: This research proposes a method to increase the accuracy of brain tumor MRI image segmentation using the ResNet model by implementing the Nearest Neighbor Upsampling method. Methods/Study design/approach: The method used is Nearest Neighbor Upsampling on ResNet to enhance image dimensions and fill gaps in MRI brain tumor images during the learning process, preserving spatial information and context crucial for segmentation. Result/Findings: The optimization of the brain segmentation model for classifying brain tumor regions using ResNet and Nearest Neighbor Upsampling achieves an increase in accuracy from 96.94% to 98.44% and a decrease in loss value from 0.0881 to 0.0874. Novelty/Originality/Value: This research addresses the limitations of the Residual Network (ResNet) model, such as the drawback of reducing the scale of images and feature maps during downsampling, resulting in a loss of spatial resolution. To overcome this challenge, the study introduces the Nearest Neighbor Upsampling method applied to ResNet, demonstrating its effectiveness in representing spatial information and image context, thereby improving segmentation accuracy.

Optimization of Residual Network 50 using Boosted Anisotropic Diffusion Filter and Contrast Limited Adaptive Histogram Equalization for Fingerprint Classification Ahmad Syafii; Riza Arifudin
Recursive Journal of Informatics Vol. 4 No. 1 (2026): March 2026
Publisher : Universitas Negeri Semarang

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.15294/rji.v4i1.13398

Abstract. Biometrics itself can be interpreted as a computerized method that uses aspects of biology, especially unique characteristics possessed by humans. Unique characteristics that can be used include fingerprints, geometric shapes of the hand, sound frequency keys, iris patterns, and retinas that generally differ from one individual to another. Fingerprints are the result of reproduction of the palm of the finger, either intentionally taken, stamped with ink, or marks left on objects because they have been touched by the skin of the palms of the hands or feet. Fingerprints are used as identification and verification as a means for security. So, there is a tool used to carry out this task, namely AFIS (Automatic Fingerprint Identification System). The purpose of this system is to strive for strong and fast detection. So, a fingerprint grouping or classification is needed, so that the identification process takes place faster. The algorithm used to classify fingerprints is ResNet-50. The data used came from the National Institute of Standards and Technology in 2000 (NIST-DB4 in 2000). In this dataset, there are 4000 data with each number per class is 800 data. There are five classes in this dataset including arch. right loop, left loop, tended arch, and whorl. In the training process, data processing is carried out first. This is done to optimize the accuracy produced during the training process. This research used preprocessing Boosted Anisotropic Diffusion Filter (BADF) and Contrast Limited Adaptive Histogram Equalization (CLAHE). The BADF method is used to reduce the noise present in the image. Whereas, CLAHE is used to adjust the contrast of the image. The accuracy produced using the two preprocessing was 94.5%. Purpose: This research aims to optimize fingerprint classification using ResNet-50 combined with Boosted Anisotropic Diffusion Filter (BADF) and Contrast Limited Adaptive Equalization (CLAHE) methods. Methods: This research uses the ResNet-50 method combined with the Boosted Anisotropic Diffusion Filter (BADF) and Contrast Limited Adaptive Histogram Equalization techniques CLAHE) methods. Result: This research has four experiments, including an experiment using the ResNet-50 model without using preprocessing to obtain an accuracy of 92.5%. When BADF preprocessing was applied in the data training process, the accuracy increased to 93.5%. Meanwhile, the experiment using the ResNet-50 model using preprocessing obtained an accuracy of 94%. This accuracy can still be improved by combining BADF and CLAHE preprocessing which gets an accuracy of 94.5%. Novelty: This research uses the ResNet-50 model with a preprocessing method that is combined to obtain higher accuracy. The update in this research is to apply the BADF and CLAHE methods as image preprocessing. The BADF method aims to reduce the noise present in the scattered image, while the CLAHE method is used to adjust the contrast in the image itself.

Optimization Of K-Nearest Neighbor Algorithm Using Information Gain And Hyperparameter Tuning In Adult Male Fertility Classification Muhammad Zaenal Muttaqin; Anggyi Trisnawan Putra
Recursive Journal of Informatics Vol. 4 No. 1 (2026): March 2026
Publisher : Universitas Negeri Semarang

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.15294/rji.v4i1.14868

Abstract. Male fertility plays an important role in reproductive capability and global population dynamics. Male infertility can be caused by lifestyle, health conditions, and sperm quality. This research develops a male fertility classification model with an optimized K-Nearest Neighbor (KNN) algorithm using Information Gain feature selection and hyperparameter tuning with GridSearchCV. The main problems encountered are low accuracy in prediction and high computational complexity due to many irrelevant features. To overcome this, feature selection and hyperparameter optimization methods were used. The dataset used in this research comes from the UCI Machine Learning Repository, consisting of 100 data with 10 attributes. The KNN algorithm was chosen for its simplicity and ability to classify data with multiple classes and uneven distribution. However, its accuracy is highly dependent on the proper selection of features and parameters. The Information Gain method is used for selection of significant features against the target variable, reducing model complexity and computation time. Hyperparameter tuning is performed using GridSearchCV to find the best combination of parameters. The results showed that the application of Information Gain and GridSearchCV successfully improved the classification accuracy of KNN. The final model achieved 94% accuracy, better than the previous conventional method which only reached 84%. This increase in accuracy shows that KNN optimization with this approach is effective in improving male fertility classification performance. This research is expected to contribute to the development of male fertility diagnostic technology and the implementation of more accurate prediction models in clinical practice. Purpose: The proposed model is a development based on previous research that focuses on developing the K-Nearest Neighbor algorithm with a model accuracy of 84%. this study uses feature selection techniques and hyperparameter tuning in the K-Nearest Neighbor (KNN) algorithm to improve the accuracy of the male fertility classification model. Methods/Study design/approach: To improve the curation of the male fertility classification model and to optimize the model from previous research, this study uses the feature selection technique and hyperparameter tuning technique. For this technique, 2 types of optimization are carried out, namely feature selection using Information Gain and GridSearchCV hyperparameter tuning to get the best parameter combination for the proposed model. The fertility dataset has also been used in previous studies, used in this study. Result/Findings: The proposed model obtained a high accuracy of 94%, which surpassed the model in the previous study which had an accuracy of 85% for the classification of fertility levels in men. Novelty/Originality/Value: The novelty in this research is the addition of hyperparameter tuning techniques to optimize and obtain optimal parameters in the fertility classification model. This research also aims to improve and increase the accuracy of the previous model.

Implementation of Content-Based Filtering in Book Recommender Systems Using K-Nearest Neighbor Model with Singular Value Decomposition and Word2Vec Naufal Afif Sadewa; Subhan Subhan
Recursive Journal of Informatics Vol. 4 No. 1 (2026): March 2026
Publisher : Universitas Negeri Semarang

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.15294/rji.v4i1.21281

Abstract. Books are a medium for understanding various topics, such as science, history, and culture. With the development of digital technology, accessing books has become easier, but choosing the right book among thousands of choices is a challenge. Book recommender systems are an efficient solution to help users find relevant books. One approach that can be used in book recommender systems is Content-Based Filtering, which utilizes content information in books to provide recommendations. Purpose: This research aims to develop a book recommender system by implementing Content-Based Filtering using K-Nearest Neighbor model with a combination of Singular Value Decomposition and Word2Vec in recommending relevant books according to each preference. Methods/Study design/approach: The method used involves several stages. First, data preprocessing is carried out to remove noise so as to retain important information. After data preprocessing, feature extraction uses the Term Frequency-Inverse Document Frequency method to represent book features through vectors. The result of this vector is then reduced in dimension using Singular Value Decomposition to reduce complexity and capture the most significant data structures. At another stage, book features are extracted using Word2Vec, which produces a semantic representation of the word in vector form. Next, the vector results from Singular Value Decomposition and Word2Vec are combined to form more informative features. Finally, the K-Nearest Neighbor model using the cosine similarity distance metric is used to calculate the similarity between books based on the combined features, so as to generate relevant book recommendations. Tests were conducted on the GoodReads Best Books dataset taken from Kaggle. Result/Findings: The evaluation results show that the proposed model is able to provide recommendations with good relevance values, measured using evaluation metrics such as Mean Average Precision and Normalized Discounted Cumulative Gain with the scores obtained respectively, namely 0.9637 and 0.9515 at parameter is 5. Novelty/Originality/Value: The novelty in this research is the combination of Singular Value Decomposition vectors and Word2Vec vectors to produce a more informative feature representation, by utilizing statistical relationships between words and capturing the semantic meaning of words.

Comparative Analysis of BERT, RoBERTa and ALBERT Model Performance with Text Data Augmentation in Multilabel Toxic Comment Classification Annisa Kunarji Sari; Zaenal Abidin
Recursive Journal of Informatics Vol. 4 No. 1 (2026): March 2026
Publisher : Universitas Negeri Semarang

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.15294/rji.v4i1.25436

Abstract. Toxic comments on social media pose serious challenges to online safety and moderation efforts. These comments are often multilabel in nature and suffer from class imbalance, making them difficult to classify accurately using standard methods. Purpose: This study investigates the use of three transformer-based language models, BERT, RoBERTa, and ALBERT, for multilabel toxic comment classification through fine-tuning. The main objective is to address class imbalance and evaluate model performance after data augmentation. Methods/Study design/approach: The Toxic Comment Classification dataset, consisting of six overlapping labels, was used in this study. A data augmentation strategy was applied using synonym replacement techniques from WordNet and easy data augmentation (EDA) to increase the representation of minority classes. After balancing the data, the dataset was split into training, validation, and testing sets. Each transformer model was fine-tuned using the Hugging Face Transformers library with the same hyperparameter settings. Model evaluation was conducted using accuracy, precision, recall, and both micro and macro F1-scores. Result/Findings: The RoBERTa model achieved the best performance, with 86.73% accuracy and a micro F1-score of 92.35%, outperforming BERT and ALBERT. The macro F1-score also improved significantly compared to previous studies using imbalanced datasets, indicating better recognition of minority classes such as threat and identity hate. Novelty/Originality/Value: This study highlights the effectiveness of combining text data augmentation with transformer-based models in handling multilabel classification tasks involving imbalanced data. The use of simple augmentation methods notably improves performance and fairness across labels, contributing to the development of more robust toxic comment detection systems.

Classification of Fresh Salmon Fish Based on Ensemble LearningUsing ResNet50 and EfficientNetV2 Arko Dwiantoro; Abas Setiawan
Recursive Journal of Informatics Vol. 4 No. 1 (2026): March 2026
Publisher : Universitas Negeri Semarang

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.15294/rji.v4i1.26230

Abstract. The increasing demand for fresh salmon in Indonesia, despite it not being a producing country, poses challenges in maintaining product quality during distribution. Freshness is a critical factor due to the fish's high susceptibility to spoilage, which can lead to health risks and economic losses. Purpose: Traditional inspection methods are inefficient for large-scale operations. Therefore, this study aims to develop an efficient and accurate classification model for fresh and infected salmon using ensemble learning based on Convolutional Neural Networks (CNN), particularly ResNet50 and EfficientNetV2 architectures. Methods/Study design/approach: This research employs a quantitative approach using the SalmonScan dataset, consisting of 1,208 images divided into two classes: fresh and infected salmon. The data underwent preprocessing, including resizing and normalization. Two deep learning architectures, ResNet50 and EfficientNetV2, were applied using the transfer learning method. These models were then combined using ensemble learning with a concatenation strategy to enhance performance. Model evaluation was conducted using accuracy, precision, recall, and F1-score, based on the confusion matrix. Results/Findings: Individual testing of ResNet50 and EfficientNetV2 models achieved high performance, but the ensemble of both architectures yielded the best results. The combined model achieved an accuracy of 98.33%, outperforming other models used in the experiment. These results indicate that the ensemble approach successfully improves the model's capability to classify salmon freshness and infection conditions. Novelty/Originality/Value: This study presents a novel ensemble approach that integrates ResNet50 and EfficientNetV2 for classifying salmon freshness. Unlike previous works that utilized either single models or more computationally expensive ensemble methods with multiple architectures, this study provides a balanced, computationally efficient solution with high accuracy. The proposed method demonstrates potential for scalable applications in fish quality assessment systems, supporting food safety and sustainability in the fisheries industry.

Optimization of Brain Tumor Segmentation on Magnetic Resonance Imaging (MRI) Using Attention Gate U-Net Fariska Ratna Fauziah; Budi Prasetiyo
Recursive Journal of Informatics Vol. 4 No. 1 (2026): March 2026
Publisher : Universitas Negeri Semarang

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.15294/rji.v4i1.27152

Abstract. Brain tumor segmentation using Magnetic Resonance Imaging (MRI) plays a vital role in medical diagnosis, requiring high precision to support clinical decisions and reduce mortality rates. Purpose: This research aims to enhance the segmentation process by implementing an Attention Gate into the U-Net model. Methods/Study design/approach: In the segmentation stage, Attention Gate on U-Net is integrated to filter out relevant information from the extracted features, resulting in a more precise segmentation of the brain tumor to determine the location of the tumor. Result/Findings: The performance of the model is assessed by calculating several evaluation metrics such as dice coefficient and intersection-over-union (IoU) for the segmentation process. The results showed that adding Attention Gate to the U-Net achieved a dice coefficient of 87.08% and IoU of 72.70% Novelty/Originality/Value: The novelty of this study lies in the integration of the Attention Gate mechanism within the U-Net decoder stage to enhance focus on tumor regions. While U-Net is widely used in medical image segmentation, this specific attention-based enhancement significantly improves performance compared to conventional U-Net models without attention. This research contributes to advancing more accurate and efficient decision-support systems in the field of medical image analysis.

XGBoost Implementation for Marine Fish Image Classification Using VGG19-Based Transfer Learning Anggito Wahyu Adi; Florentina Yuni Arini
Recursive Journal of Informatics Vol. 4 No. 2 (2026): September 2026
Publisher : Universitas Negeri Semarang

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.15294/rji.v4i2.37834

Abstract. Accurate identification of fish species is important for trade supervision, marine biology, and fisheries management. This study develops a classification system for nine marine fish species by combining VGG19-based feature extraction with XGBoost. Using 9,000 balanced RGB images from the Kaggle Large-Scale Fish Dataset, the workflow includes EDA, preprocessing, an 80:20 data split, feature extraction, and model training. EDA showed that Red Mullet and Striped Red Mullet share similar visual characteristics, while the Shrimp class appears more distinct. The model achieved 97% accuracy and was deployed through a Flask web application that processes, extracts features, and classifies images. These findings demonstrate that integrating VGG19 with XGBoost offers an effective and efficient approach for marine fish image classification. Purpose: The primary objective here is to build a hybrid classification model for marine fish that is both efficient and robust, merging deep feature extraction from the VGG19 architecture with the XGBoost classifier. Specifically, this research seeks to improve classification accuracy while reducing computational overhead, thereby providing a reliable automated identification system that can support fisheries management, ecological monitoring, and marine biological research. Methods: This research employed a dataset of 9,000 images representing nine marine fish species. The preprocessing stage included resizing each image to 224×224 pixels and applying normalization. Deep feature representations were obtained using a pre-trained VGG19 model, utilizing only its convolutional layers to extract semantic features. These features were then classified with the XGBoost algorithm, chosen for its efficiency, resilience, and strong performance on structured data. The model was trained using an 80:20 split for training and testing, and its performance was assessed through accuracy, precision, recall, F1-score, and a confusion matrix. Result/Findings: The combination of VGG19 feature extraction with XGBoost produced strong results, achieving an accuracy of about 97%. This performance was further supported by consistently high precision, recall, and F1-scores across all nine species. Novelty/Value: A hybrid architecture integrating VGG19-based transfer learning with the XGBoost classifier is presented with an approach rarely explored in this domain. This combination capitalizes on deep semantic feature extraction from VGG19 alongside efficient structured classification via XGBoost, producing a model that delivers high accuracy with significantly reduced computational costs. Additionally, the hybrid approach proposed in this study remains competitive with fully end-to-end deep learning models yet operates with fewer parameters and reduced training demands. These results illustrate that well-designed hybrid pipelines can match the performance of conventional deep learning architecture while providing clear benefits in terms of training efficiency, interpretability, and scalability.

Sentiment Analysis of Generative AI Application Reviews Using Random Forest and SVM with SMOTE-Based Class Imbalance Handling Istiqomah; Alamsyah
Recursive Journal of Informatics Vol. 4 No. 2 (2026): September 2026
Publisher : Universitas Negeri Semarang

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.15294/rji.v4i2.60530

Abstract. The rapid adoption of Generative Artificial Intelligence (Gen-AI) has generated extensive user reviews, offering valuable insights into user perceptions. This study aims to classify review sentiment using Random Forest and Support Vector Machine (SVM), evaluate the impact of SMOTE on class imbalance, and analyze review themes. This study analyzed 50,000 reviews from five popular Gen-AI applications: ChatGPT, Microsoft Copilot, Google Gemini, Perplexity, and Claude. The models were evaluated using accuracy, weighted precision, weighted recall, weighted F1-score, and confusion matrix. The results showed that SVM without SMOTE achieved the highest overall accuracy of 84% and a weighted F1-score of 0.81. However, its F1-score for the neutral class was only 0.06. Applying SMOTE reduced the overall accuracy of Random Forest from 83% to 81% and SVM from 84% to 76%. Nevertheless, SMOTE increased the neutral-class F1-score from 0.00 to 0.06 for Random Forest and from 0.06 to 0.20 for SVM. These results indicate that SMOTE improved minority-class recognition, particularly for SVM, but introduced a trade-off with overall predictive performance. Ultimately, these findings provide practical insights for improving Gen-AI applications while also serving as a reference for selecting sentiment classification models in imbalanced multiclass sentiment analysis. Purpose: This study aims to classify the sentiment of user reviews on Gen-AI applications using Random Forest and Support Vector Machine (SVM), evaluate the impact of SMOTE on imbalanced multiclass data, and identify key review themes influencing user perceptions. Methods/Study design/approach: A dataset of 50,000 user reviews from five popular Gen-AI applications was analyzed using text preprocessing, TF-IDF feature extraction, rating-based sentiment labeling, and SMOTE applied to the training data. Random Forest and SVM were evaluated using accuracy, weighted precision, weighted recall, weighted F1-score, and confusion matrix. Result/Findings: SVM without SMOTE achieved the best overall performance, with an accuracy of 84% and a weighted F1-score of 0.81. Although SMOTE reduced overall accuracy, it improved the neutral-class F1-score from 0.00 to 0.06 for Random Forest and from 0.06 to 0.20 for SVM, demonstrating better recognition of minority-class samples. Novelty/Originality/Value: This study provides a comparative evaluation of Random Forest and SVM with and without SMOTE for multiclass sentiment analysis of Gen-AI user reviews. It also integrates review theme analysis to identify factors affecting user perceptions, offering practical guidance for selecting classification models and improving Gen-AI applications.

Comparative Analysis of Decision-Level Ensemble Voting Strategies for Tomato Leaf Disease Classification Using EfficientNet-B0 and MobileNetV2 Hani Nafis Muzaki; Yahya Nur Ifriza
Recursive Journal of Informatics Vol. 4 No. 2 (2026): September 2026
Publisher : Universitas Negeri Semarang

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.15294/rji.v4i2.61961

Abstract. Tomato leaf diseases are a major threat to agricultural productivity, making accurate and timely disease classification essential for supporting precision agriculture. Recent advances in convolutional neural networks have significantly improved automated plant disease recognition, while ensemble learning has further enhanced classification performance by combining predictions from multiple models. Among various ensemble approaches, decision-level voting strategies, including hard voting, soft voting, and weighted soft voting, are widely adopted because of their simplicity and effectiveness. However, systematic comparisons of these voting strategies, particularly when combining lightweight CNN architectures for tomato leaf disease classification, remain limited. Therefore, this study investigates the comparative performance of different decision-level voting strategies to provide empirical evidence for selecting the most appropriate ensemble approach. Purpose: This study aims to systematically compare hard voting, soft voting, and weighted soft voting for combining EfficientNet-B0 and MobileNetV2 to identify the most effective decision-level ensemble strategy for tomato leaf disease classification. Methods/Study design/approach: EfficientNet-B0 and MobileNetV2 were trained independently and combined using hard voting with confidence-based tie breaking, soft voting, and weighted soft voting, with voting weights optimized through grid search. Performance was evaluated using accuracy, macro-F1 score, confusion matrices, and McNemar's test. Result/Findings: EfficientNet-B0 outperformed MobileNetV2 as the best single model (94.05% accuracy, 0.9266 macro-F1 vs. 90.64% accuracy, 0.8781 macro-F1). All three ensemble strategies further improved classification performance, with weighted soft voting achieving the highest performance (95.15% accuracy, 0.9417 macro-F1). McNemar's test showed that only soft voting and weighted soft voting significantly outperformed the best single model, whereas hard voting did not. No statistically significant differences were observed among the three ensemble voting strategies. Novelty: This study systematically compares hard voting, soft voting, and weighted soft voting for decision-level ensemble using EfficientNet-B0 and MobileNetV2 in tomato leaf disease classification. In addition to evaluating predictive performance, the study incorporates McNemar's statistical test to determine whether the observed performance differences among the ensemble strategies are statistically significant, providing empirical guidance for selecting appropriate voting strategies.

Home Page

OAI Link

Editorial Team

Contact

Reviewer

Google Scholar

Contact Name
-

Contact Email
rji@mail.unnes.ac.id

Phone
-

Journal Mail Official
rji@mail.unnes.ac.id

Editorial Address
Sekaran, Kec. Gn. Pati, Kota Semarang, Jawa Tengah 50229

Location
Kota semarang,

Jawa tengah

INDONESIA

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Home Page

OAI Link

Editorial Team

Contact

Reviewer

Google Scholar

Contact Name -

Contact Email rji@mail.unnes.ac.id

Phone -

Journal Mail Official rji@mail.unnes.ac.id

Editorial Address Sekaran, Kec. Gn. Pati, Kota Semarang, Jawa Tengah 50229

Location Kota semarang, Jawa tengah INDONESIA

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Contact Name
-

Contact Email
rji@mail.unnes.ac.id

Phone
-

Journal Mail Official
rji@mail.unnes.ac.id

Editorial Address
Sekaran, Kec. Gn. Pati, Kota Semarang, Jawa Tengah 50229

Location
Kota semarang,

Jawa tengah

INDONESIA