Jurnal Informatika: Jurnal Pengembangan IT
Not a member yet
437 research outputs found
Sort by
Monk Skin Tone Classification: RMSprop vs Adam Optimizer in MobileNetV2
No full textThe lack of accurate and accessible skin tone classification systems poses significant challenges in personalized fashion recommendations and inclusive technology development. This study aims to develop a skin tone classification system utilizing the Monk Skin Tone (MST) scale through the implementation of Convolutional Neural Network with MobileNetV2 architecture enhanced by transfer learning techniques. The MST scale encompasses ten distinct categories providing comprehensive representation of human skin color diversity. The methodology leverages efficient MobileNetV2 architecture suitable for web deployment, transfer learning to enhance accuracy despite limited training data, and strategic dataset balancing. A dataset of 1,729 facial photographs representing the complete MST spectrum was utilized. Preprocessing involved scaling images to 224×224 pixels, normalization, and augmentation through various transformations to address class imbalance challenges. The dataset was partitioned using a 70:15:15 ratio for training, validation, and testing respectively. The system was implemented as a web platform called SkinToneAI that enables users to upload facial images for skin tone analysis and receive personalized clothing color recommendations. Evaluation demonstrated classification accuracy of 97.83% on the test dataset with a loss value of 0.1166 when using Adam optimizer, while RMSprop optimizer achieved better performance with 98.26% accuracy and 0.0548 loss value. The implemented web application successfully translates technical capabilities into practical fashion assistance. The system provides users with customized apparel color suggestions based on their identified skin tone category, effectively connecting advanced AI technology with everyday fashion needs
Dinamika Opini Publik Terkait Quarter Life Crisis Pada Media Sosisal X Menggunakan Support Vector Machine
No full textThis study aims to analyze the dynamics of public opinion related to quarter life crisis on platform X through a sentiment analysis approach based on machine learning Support Vector Machine (SVM) algorithm is used to classify positive and negative sentiments from text data. A total of 6.312 tweets were collected with the keyword “quarter life crisis” from January 2024 to January 2025. The data was then processed through the stages of text cleaning, tokenization, stopword removal, stemming, and lexicon-based sentiment labeling. The classification process is carried out using SVM with a data division of 80% training and 20% test. The results showed an accuracy of 81.57% with a sentiment distribution of 59.3% negative and 40.7% positive. Implementation was done on Google Colab platform with evaluation using confusion matrix and classification report. The fingdings prove the effectiveness of SVM in analyzing psychosocial phenomena on social media and provide an empirical basis for the development of digital data-based mental health interventions. The machine learning pipeline optimized in this study can be used as a reference for other studies in analyzing psychological phenomena on social medi
Random Under Sampling for Performance Improvement in Attack Detection on Internet of Vehicles Using Machine Learning
No full textThe Internet of Vehicles (IoV) technology is one of the advancements derived from the Internet of Things (IoT) in the transportation sector, benefiting its users. However, the development of this technology cannot be separated from various security threats, particularly Denial of Service (DoS) and spoofing attacks. Given these threats, it is crucial to continuously develop methods used for detecting attacks on IoV systems. Several researchers have conducted research related to attacks and threats on IoV systems, and one such study resulted in a dataset called CICIoV2024. This dataset has an imbalanced class distribution. This study aims to examine the implementation of Random Under-Sampling to improve the performance of classification algorithms in detecting attacks on IoV systems. The algorithms used in this study include Decision Tree, K-Nearest Neighbors (KNN), and Random Forest. The test results show that the Random Forest algorithm achieved the best results with an accuracy of 98.5% and an F1-Score of 98.5%
Enhancing PCOS Classification with Weighted Loss-Based Neural Network on Imbalanced Data
No full textPolycystic Ovary Syndrome (PCOS) represents a multifaceted endocrinemetabolic condition that poses a significant risk to reproductive health in women of childbearing age. The disorder is influenced by various contributing factors and is commonly associated with clinical features such as disrupted ovulation, hormonal imbalance due to excess androgens, and morphological changes in the ovaries. In automated PCOS classification, a major limitation arises from the disproportionate distribution of data samples, in which instances without PCOS considerably outnumber affected cases. This imbalance tends to bias predictive models toward the dominant class, thereby reducing the detection capability for minority instances and increasing the likelihood of missed PCOS diagnoses. To address this issue, this study proposes the incorporation of a Weighted Loss Function into a Neural Network-based classification framework aimed at improving sensitivity to PCOS cases. The research workflow comprises data preprocessing, neural network architecture construction, integration of class-weighted loss, and systematic experimentation across multiple architectural designs and training configurations. The experimental findings demonstrate that applying a Weighted Loss Function with manually assigned class weights of 1:2, a learning rate of 0.001, five hidden layers, and 50 training epochs delivers optimal classification performance. Under these settings, the model achieves high values across evaluation metrics, including precision, recall, F1-score, and overall accuracy, reaching up to 99%. The results confirm that the proposed approach effectively mitigates majority-class bias and enhances the models ability to identify PCOS cases. This improvement is further reinforced through careful hyperparameter tuning and comprehensive experimental evaluation
Segmentasi Teks Arab Pegon Menggunakan Histogram Segmentation
No full textAksara Pegon merupakan salah satu warisan budaya Nusantara yang memiliki nilai sejarah tinggi, terutama dalam literatur Islam di Jawa, Sunda dan Madura. Upaya digitalisasi naskah Pegon menghadapi sejumlah kendala yang tidak sederhana. Kompleksitas bentuk huruf, pola penyambungan antarhuruf, serta ketidakteraturan tata letak tulisan pada manuskrip kuno menjadikan proses ekstraksi teks jauh lebih menantang dibandingkan teks cetak modern. Salah satu tahapan yang sangat penting dalam proses tersebut adalah segmentasi teks. Ketepatan pemisahan karakter atau kata akan sangat menentukan keberhasilan tahap berikutnya, yaitu Optical Character Recognition (OCR) dan transliterasi otomatis. Histogram segmentation dikenal mampu mendeteksi dan memisahkan objek teks dari latar belakang dengan memanfaatkan distribusi intensitas piksel. Metode ini sudah dikenal luas dalam segmentasi citra digital. Cara kerjanya yang sederhana tanpa memerlukan pelatihan, membuat metode ini cocok digunakan sebagai langkah awal menuju tahap transliterasi aksara Arab-Pegon. Hasil eksperimen menunjukkan bahwa model mampu melakukan segmentasi baris dengan baik. Sedangkan tingkat akurasi segmentasi kata sebesar 70% dengan error rate Over-Segmentation Rate (OSR) sebesar 0.20 dan Under-Segmentation Rate (USR) sebesar 0.13. Metode histogram segmentation terbukti ringan, cukup efisien, dan tidak memerlukan pelatihan data seperti pada pendekatan deep learning. Penelitian ini memiliki kontribusi besar terhadap pelestarian aksara Arab-Pegon sebagai khazanah ilmu pengetahuan yang dimiliki oleh nusantara
Pengembangan Aplikasi Prediksi Harga Emas Berbasis Web Menggunakan Model Time Series
No full textHigh gold price volatility due to global economic instability poses challenges in investment decision-making. This research aims to develop a web-based gold price prediction application using a time series model, focusing on the Gated Recurrent Unit (GRU) algorithm. This application is designed to present real-time, accurate, and easily accessible gold price predictions, thereby increasing the efficiency and transparency of information for investment decision making. The development process starts from collecting and preprocessing daily gold price data for the period 2013-2023, then comparing four predictive models: LSTM, GRU, ARIMA, and XGBoost. Evaluation is performed using MAE, RMSE, and R² metrics. Results showed that GRU provided the best performance with an RMSE value of 17.76 and R² of 0.9410. The GRU model is integrated into a web application using the Flask framework, with an interactive HTML-based interface and Chart.js visualization. This application presents real-time gold price predictions and can be accessed by general users and investors. The results of this study show that the time series approach with GRU is effective in projecting gold prices, and can be a relevant tool in supporting data-based investment decisions
Klasifikasi Pertanyaan Quora Menggunakan Metode Keyword-based dan Analisis Sentimen dengan ComplementNB
No full textText classification is a fundamental task in Natural Language Processing (NLP) that supports the categorization of data based on predefined labels. This study aims to evaluate the effectiveness of keyword-based labeling and sentiment analysis methods for text classification using the Quora Questions dataset. The dataset comprises 16,921 samples with imbalanced class distribution, where the opinion category dominates, while the hypothetical category is a minority class. The labeling process utilized a keyword-based approach for the fact and hypothetical categories, while the opinion category was labeled using sentiment analysis with the Vader Lexicon library. TF-IDF was employed as the feature representation method, with two approaches explored: n-gram range tuning (1–3) and without tuning. ComplementNB, designed for handling imbalanced datasets, was utilized for classification, with a training-test split of 70:30. The results show that the approach without n-gram tuning achieved the highest accuracy of 93.89%, with zero variance in cross-validation. Evaluation revealed that ComplementNB effectively handles class imbalance, as demonstrated by high precision and recall in the minority class. This study demonstrates that a simple approach combining keyword-based labeling and sentiment analysis can be effectively implemented for category-based text classification tasks, particularly in platforms like Quora. These findings are relevant for similar applications requiring real-time text classification with minimal complexity
Pemodelan Topik pada Komentar YouTube Arra: Komparasi LDA dan K-Means Menggunakan Fitur Leksikal dan Semantik
No full textYouTube has become a platform for sharing content, including positive material and stereotypes that often trigger debates. One noteworthy phenomenon is the video of Arra, a toddler known for her remarkable communication skills. This uniqueness has drawn significant attention and sparked debates about the mismatch between her age and cognitive development. The diverse comments on Arra’s videos reflect sharply differing perspectives among netizens, making manual analysis highly challenging. Therefore, it is important to examine the topics discussed by netizens to understand the dominant issues emerging in these discussions. Through this approach, the public can gain insights, and parents may receive valuable input regarding child-rearing practices. The main objective of this study is to explore the effectiveness of the two methods and their combinations of text representations in identifying key topics within comments by comparing the coherence performance of the models. This research applies topic modeling to analyze comments using two primary approaches: Latent Dirichlet Allocation (LDA) and K-Means clustering. The study involves data collection through comment crawling, followed by text preprocessing and text representation using TF-IDF and GloVe embeddings. LDA and K-Means are then used to identify dominant topics appearing in the comments. The results show that LDA with TF-IDF achieved the highest coherence score of 0.662, although the resulting topics were still difficult to interpret due to overlap. Meanwhile, K-Means with GloVe 100D yielded a slightly lower coherence score of 0.6538 but outperformed in terms of interpretability. Therefore, K-Means with GloVe 100D is considered a more balanced approach in terms of both coherence and topic readability
Developing a Web-based MSME Sales Revenue Data Management and Reporting Portal Using OAuth 2.0
No full textThe Cooperatives and SMEs Service (DinKopUKM) under the Ministry of Cooperatives and SMEs play a vital role in coordinating the implementation of tasks, coaching and providing administrative support for MSMEs. In this research, a portal for managing MSME data and Sales Revenue reporting was carried out to support transparent information management, monitoring MSME business implementation, and encouraging orderly administration in MSMEs so as to optimize efforts to empower MSMEs. The portal was developed using Laravel technology for the backend and NextJS for the frontend, with a responsive web design so that it can be accessed from various devices. Apart from that, to support data integration and data communication with resources that have been built in previous research, OAuth 2.0 was implemented. The development process is in accordance with the Prototyping process model. The portal developed was tested using the black box testing method. It was found that this portal was in accordance with the needs and design of the system. The portal developed helps DinKopUKM in carrying out its duties of data collection on MSMEs, monitoring MSMEs, and encouraging MSMEs to maintain orderly administratio
Evaluasi User Experience Google Lens pada Pengguna Baru Menggunakan Metode Cognitive Walkthrough
No full textNew users often experience difficulties using Google Lens due to a lack of familiarity with Augmented Reality (AR) technology. This study aims to evaluate the usability of Google Lens using the Cognitive Walkthrough method, focusing on three main task scenarios: text translation, object identification, and barcode scanning. A total of 60 respondents participated in the study, with data collected through a Likert-based questionnaire (1–5) and direct observation of user interactions. The quantitative analysis results showed that Google Lens obtained an average score of 3.5 for satisfaction, 3.8 for efficiency, and 3.2 for ease of navigation. Evaluation per scenario showed the highest success rate for barcode scanning (92%), followed by text translation (85%), while object identification had the lowest success rate (78%). Qualitative findings revealed that less intuitive navigation, unclear function icons, a lack of initial guidance, and limited object identification accuracy were the main obstacles for new users. Based on these results, this study recommends several improvements, including optimizing the interface design, adding descriptive labels to icons, providing personalization features, and developing interactive tutorials for new users. With these recommendations, it is hoped that Google Lens can become an application that is easier to learn, more efficient, and provides a more satisfying experience, while also enriching the literature related to usability evaluation of Augmented Reality-based applications.