IJATIS: Indonesian Journal of Applied Technology and Innovation Science

Application of Deep Learning for the Classification of Brain Tumor Magnetic Resonance Imaging Images

Rizki Andreas, Mustakim Mustakim, Ma. Angelica M. Qin, Safril Siregar — Tue, 28 Jul 2026 00:00:00 +0000

Accurate and timely classification of brain tumors from Magnetic Resonance Imaging (MRI) is essential for supporting clinical diagnosis and treatment planning. This study presents a comparative evaluation of transfer learning-based Convolutional Neural Network (CNN) architectures, including ResNet-50, DenseNet121, and EfficientNet-B2, for binary brain tumor MRI classification (Glioma and Meningioma). The dataset was augmented to improve model generalization, and the data were split into training and test sets using an 80:20 hold-out split. The models were trained using Adam and RMSProp optimizers with different learning rates, and their performance was evaluated using accuracy, precision, recall, and F1-score. Experimental results demonstrate that data augmentation significantly enhances classification performance across all evaluated architectures. Among the tested models, ResNet-50 with the RMSProp optimizer and a learning rate of 0.001 achieved the best performance, yielding the lowest training and validation losses and perfect classification results on the test set, with accuracy, precision, recall, and F1-score approaching 100%. DenseNet121 and EfficientNet-B2 also achieved excellent performance but were slightly inferior to ResNet-50. These findings indicate that transfer learning with ResNet-50 and RMSProp provides an effective and reliable solution for automated brain tumor MRI image classification

Effective Machine Learning Schemes for Face Recognition in Higher-Dimensional Datasets

Onibeju Basit Adeyinka, Akinyemi Moruff Oyelakin — Fri, 31 Jul 2026 00:00:00 +0000

Face recognition systems play a vital role in personal identification, particularly in security applications. However, high-dimensional image data often reduces the effectiveness of machine learning (ML) models by increasing computational complexity and training difficulty. This study aims to improve face recognition performance by combining dimensionality reduction and deep feature extraction techniques. Two benchmark datasets, Filtered LFW and a custom VGGFace2 dataset, were used to evaluate two ML-based models. The first model integrates Linear Discriminant Analysis (LDA), CNN (ResNet18), and K-Nearest Neighbors (KNN), while the second combines CNN (ResNet18) with Naïve Bayes (NB). Both models were trained, tested, and evaluated using accuracy, precision, recall, F1-score, and execution time. On the LFW dataset, the CNN-LDA-KNN model achieved the best performance, with 97.82% accuracy, 0.9799 precision, 0.9782 recall, 0.9782 F1-score, and 68.85 seconds evaluation time, outperforming the CNN-NB model (96.71% accuracy). Similarly, on the VGGFace2 dataset, CNN-LDA-KNN obtained 93.46% accuracy compared with 78.87% for CNN-NB. These findings demonstrate that integrating LDA with CNN-based feature extraction and KNN classification significantly enhances face recognition performance on high-dimensional image datasets while maintaining competitive computational efficiency.

Social Media Sentiment Analyst on Danantara Using Support Vector Machine with TF-IDF Weighting

Wed, 30 Sep 2026 00:00:00 +0000

The establishment of Badan Pengelola Investasi Daya Anagata Nusantara (Danantara) has generated extensive public discussion on social media, reflecting diverse perceptions regarding its governance, transparency, and potential contribution to national development. Despite growing public discourse, empirical studies examining public sentiment toward this newly established institution remain limited. This study addresses that gap by analyzing public sentiment toward Danantara using a Support Vector Machine (SVM) classifier with Term Frequency–Inverse Document Frequency (TF-IDF) feature weighting, applied to a multi-platform dataset spanning X (Twitter), Instagram, and YouTube an approach that captures a broader and more representative range of public opinion than single-platform studies. A dataset of 3,000 comments was collected and underwent preprocessing, including case folding, cleaning, normalization, tokenization, stopword removal, and stemming, followed by manual labeling into three sentiment classes: positive, negative, and neutral. The proposed LinearSVC model achieved an accuracy of 84.67%, a macro-average F1-score of 0.82, and a weighted-average F1-score of 0.85. Sentiment distribution analysis reveals that negative sentiment dominates public discourse (57.13%), followed by positive (24.37%) and neutral sentiments (18.50%). This study contributes both a novel empirical case being among the first sentiment analyses of public perception toward Danantara and evidence that SVM with TF-IDF offers an effective, robust approach for multiclass sentiment classification of Indonesian social media text, providing actionable insights for policymakers and institutional stakeholders.

NLP-Based Analysis of Indonesian Public Opinion on the Nadiem Makarim Chromebook Case Using Instagram and YouTube Comments

Wed, 30 Sep 2026 00:00:00 +0000

This study examines Indonesian public sentiment toward the Ministry of Education, Culture, Research, and Technology (Kemendikbudristek)'s Chromebook procurement under Nadiem Makarim and evaluates emoji tokenization in text classification. Between 2021 and 2024, the initiative sparked debate over device specifications, internet connectivity in frontier, outermost, and underdeveloped (3T) regions, and budget accountability. A balanced dataset of 4,794 comments (2,397 from Instagram and 2,397 from YouTube) was analyzed after removing duplicates. Sentiment pseudo-labels (negative, positive, and neutral) were generated using Indonesian RoBERTa fine-tuned on the SmSA IndoNLU benchmark. Two preprocessing scenarios were compared: removing emojis and emoticons (E1) and converting them into descriptive tokens (E2). Unigram and bigram TF-IDF features were classified using class-weighted Multinomial Logistic Regression within a leakage-free pipeline. Negative sentiment dominated (76.93%; 3,688 comments), followed by positive (17.86%; 856 comments) and neutral (5.21%; 250 comments). In the 80:20 holdout evaluation, E2 achieved 0.8352 accuracy and 0.6722 macro F1, while E1 achieved 0.8321 accuracy and 0.6725 macro F1. Five-fold cross-validation showed modest improvements with E2 in mean accuracy (0.8204 vs. 0.8169) and macro F1 (0.6433 vs. 0.6333). Performance was higher on Instagram than on YouTube. Overall, findings indicate predominantly negative public reception and modest benefits from emoji tokenization.

Evaluating the Role of PCA Dimensionality Reduction in Enhancing CNN-Based Butterfly Image Classification

Mon, 05 Oct 2026 00:00:00 +0000

Butterfly species classification from digital images is a challenging computer vision task due to high intra-class variability across 75 species. Although Convolutional Neural Networks (CNN) have demonstrated strong performance in biological image classification, high-dimensional image data raises concerns about computational cost and the risk of overfitting. The main novelty of this study is a rigorous empirical evaluation of Principal Component Analysis (PCA) as a dimensionality-reduction preprocessing step in a CNN-based butterfly classification pipeline, a combination that has not been explicitly investigated in this domain. Three comparative experimental scenarios were designed: Scenario A used VGG16 transfer learning without PCA as a baseline; Scenario B applied PCA retaining 95% explained variance before a Multilayer Perceptron (MLP) classifier; and Scenario C applied PCA retaining 99% explained variance with the same MLP architecture. The performance evaluation covered accuracy, precision, recall, F1 Score, training time, inference time, and model size. Results show that Scenario A achieved the highest accuracy of 62.05% (F1-score: 0.6158), while Scenarios B and C reached 35.90% and 36.62% respectively, with training time reduced by up to 98.5% compared to the baseline. These findings reveal a clear accuracy-efficiency trade-off: PCA substantially reduces computational cost but at the expense of classification accuracy when applied as preprocessing before MLP rather than within CNN layers. This study offers practical empirical evidence to guide practitioners in selecting between accuracy-optimized and efficiency-optimized strategies when computational resources are constrained

Facial Expression Classification Using Convolutional Neural Network on the FER2013 Dataset

Muhammad Farrel Adynata, Verrel Muhammad Andrifan, Hadi Defri, Azmil Hasim — Mon, 05 Oct 2026 00:00:00 +0000

Facial expression classification is an important task in computer vision that enables systems to identify human emotions from facial images. This study aims to evaluate the performance of a Convolutional Neural Network (CNN) for classifying facial expressions using the FER2013 dataset. The dataset consists of grayscale facial images categorized into seven emotion classes: Angry, Disgust, Fear, Happy, Sad, Surprise, and Neutral. The images were normalized and augmented using rotation, shifting, zooming, and horizontal flipping to improve model generalization. A CNN architecture consisting of three convolutional layers with 32, 64, and 128 filters, followed by max-pooling, a fully connected layer, dropout, and a softmax output layer, was trained using the Adam optimizer for 20 epochs with a batch size of 32. Model performance was evaluated using accuracy, precision, recall, F1-score, and a confusion matrix. The model achieved a test accuracy of 60%, with weighted precision, recall, and F1-score of 60%. Among the seven classes, Happy achieved the highest F1-score of 79%, followed by Surprise at 73%, while Fear, Sad, and Disgust achieved F1-scores of 45%, 46%, and 47%, respectively. These results indicate that the CNN can effectively classify several facial expressions, particularly Happy and Surprise, while performance is lower on visually similar and less-represented classes. The contribution of this study is an evaluation and class-level analysis of a CNN-based facial expression classification approach on the FER2013 dataset, providing insights into performance differences across emotion categories and potential directions for improving classification performance.

Comparative Performance Study of Machine Learning Algorithms for Marketing Campaign Response Prediction

Sun, 04 Oct 2026 00:00:00 +0000

This study presents a comparative analysis of five Machine Learning (ML) classification algorithms, namely Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), XGBoost, and Gradient Boosting Classifier (GBC), for predicting customer responses to marketing campaigns. The main challenge in this task is the highly imbalanced distribution of customer response data, which may cause biased model predictions. Therefore, this study evaluates model performance using precision, recall, and F1-score metrics to provide a more reliable assessment. Experiments were conducted using the Customer Personality Analysis dataset obtained from Kaggle. The results show that ensemble-based algorithms achieved more consistent performance than individual classifiers. XGBoost obtained the highest F1-score of 84.85% with an 80:20 data split, followed by Random Forest with 84.72%. In contrast, SVM showed lower stability, with its F1-score decreasing from 83.59% to 78.66% on the 90:10 data split. This study provides empirical evidence on the effectiveness of ensemble learning approaches, particularly XGBoost, in handling the complexity of customer behavior and supporting more targeted, data-driven marketing strategies.