Garuda - Garba Rujukan Digital

Experimenting Diabetic Retinopathy Classification Using Retinal Images Muhammad Fermi Pasha; Mark Dhruba Sikder; Asif Rana; Maya Silvi Lidya; Ronsen Purba; Rahmat Budiarto
Data Science: Journal of Computing and Applied Informatics Vol. 5 No. 1 (2021): Data Science: Journal of Computing and Applied Informatics (JoCAI)
Publisher : Talenta Publisher

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.32734/jocai.v5.i1-5232

Along with many complications, diabetic patients have a high chance to suffer from critical level vision loss and in worst case permanent blindness due to Diabetic Retinopathy (DR). Detecting DR in the early stages is a challenge, since it has no visual indication of this disease in its preliminary stage, thus becomes an important task to accomplish in the health sector. Currently, there have been many proposed DR classifier models but there is a lot of room to improve in terms of efficiency and accuracy. Despite having strong computational power, current deep learning algorithm is not able to gain the trust of the medical experts in classifying DR. In this work, we investigate the possibility of classifying DR using deep learning with Convolutional Neural Network (CNN). We implement preprocessing combined with InceptionV3 and VGG16 models. Experimental results show that InceptionV3 outperforms VGG16. InceptionV3 model achieves an average training accuracy of 73.5 % with a validation accuracy of 68.7%. VGG16 model achieves an average training accuracy of 66.4% with a validation accuracy of 63.13%. The highest training accuracy for InceptionV3 and VGG16 is 79% and 81.2%, respectively. Overall, we achieve an accuracy of 66.6% on 52 images from 3 different classes.

Stock Price Prediction Using XCEEMDAN-Bidirectional LSTM -Spline Kelvin Chen; Ronsen Purba; Arwin Halim
Indonesian Journal of Artificial Intelligence and Data Mining Vol 5, No 1 (2022): March 2022
Publisher : Universitas Islam Negeri Sultan Syarif Kasim Riau

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.24014/ijaidm.v5i1.14424

Bidirectional Long Short Term Memory (Bidirectional LSTM) is a machine learning technique with the ability to capture data context by traversing backward data to forward data and vice versa. However, the characteristics of stock data with large fluctuations, high dimensions and non-linearity become a challenge in obtaining high stock price prediction accuracy values. The purpose of this study is to provide a solution to the problem of stock data characteristics with large fluctuations, high dimensions and non-linearity by combining the Complete Ensemble Empirical Mode Decomposition With Adaptive Noise method for exogenous features (XCEEMDAN), Bidirectional Long Short Term Memory (LSTM), and Splines. The predicted data will go through normalization and preprocessing using XCEEMDAN then the XCEEMDAN decomposition results are divided into high and low frequency signals. The bidirectional LSTM handles high frequency signals and the Spline model handles low frequency signals. The test is carried out by comparing the proposed XCEEMDAN-Bidirectional LSTM-Spline model with the XCEEMDAN-LSTM-Spline model using the same parameters and changing the noise seed randomly 50 times. The test results show that the proposed model has the smallest RMSE average value of0.787213833 while model which is compared only has the smallest RMSE average value of 0.807393567.

Deteksi Potensi Depresi dari Unggahan Media Sosial X Menggunakan IndoBERT Situmorang, Gilbert Fernando; Purba, Ronsen
Building of Informatics, Technology and Science (BITS) Vol 6 No 2 (2024): September 2024
Publisher : Forum Kerjasama Pendidikan Tinggi

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.47065/bits.v6i2.5496

Over the past few decades, mental disorders such as depression have increased and become a serious public health issue. Many affected individuals choose not to seek professional support due to social stigma. Social media platforms like X provide opportunities to study mental health on a large scale because users often share their personal experiences and emotions. However, there are challenges in understanding language patterns and context in posts, necessitating appropriate techniques and models to effectively detect potential depressions. Utilizing Natural Language Processing (NLP) techniques, this study analyzes 37,554 texts from social media posts to detect potential depressions. This study employs the IndoBERT model, an adaptation of BERT trained on Indonesian text data, to identify potential depression from social media texts. Data were collected through scrapping using negatively and positively connotated keywords, which were consulted with psychiatrists. The text pre-processing includes case folding, text cleaning, spell normalization, stopword removal and stemming. The data were then labeled using the IndoBERT emotion classification model, categorizing negative emotions as depression and positive emotions as normal. The model was trained and evaluated using accuracy, precision, recall, and F1-score metrics, with the best results showing an accuracy of 94.91%, precision of 94.91%, recall of 94.91%, and an F1-score of 94.91%. The results indicate that the IndoBERT model is effective in detecting potential depression from social media texts. However, there are limitations due to the reliance on social media posts, which may not fully reflect the users’ emotional conditions.

Optimization of Sentiment Analysis Classification of ChatGPT on Big Data Twitter in Indonesia using BERT Sinaga, Frans Mikael; Purba, Ronsen; Pipin, Sio Jurnalis; Lestari, Wulan Sri; Winardi, Sunaryo
JURNAL MEDIA INFORMATIKA BUDIDARMA Vol 8, No 3 (2024): Juli 2024
Publisher : Universitas Budi Darma

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.30865/mib.v8i3.7861

This research is grounded in the emergence of ChatGPT technology, supported by prior and similar studies. The urgency of the issue is highlighted by previous research indicating non-convergent classification outcomes in LSTM (Long Short-Term Memory) methods due to suboptimal hyperparameter settings and limitations in understanding text data within Big Data. The presence of ChatGPT technology brings both benefits and potential misuse, such as copyright infringement, unauthorized news extraction, and violations of accountability principles. Understanding public sentiment towards the presence of ChatGPT technology is crucial. The research aims to implement the BERT (Bidirectional Encoder Representations from Transformers) method to achieve accurate and convergent sentiment analysis classification. This study involves data preprocessing stages using Natural Language Processing (NLP) techniques. Text data, already vectorized, is classified using BERT to determine public sentiment (positive, negative, neutral) towards ChatGPT technology, ensuring greater accuracy, convergence, and contextual relevance. Performance testing of the BERT model is conducted using a Confusion Matrix. With parameters set to Max Sequence Length = 128 and Batch Size = 16, the highest classification accuracy achieved is 93.4%.

Movie Success Prediction Based on Feature and Trailer Comments Using Ensemble+LSTM Model Nadya Sikana; Purba, Ronsen
Journal La Multiapp Vol. 5 No. 5 (2024): Journal La Multiapp
Publisher : Newinera Publisher

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.37899/journallamultiapp.v5i5.1417

Predicting the success of a movie is a very important aspect due to the high risks involved in movie production. The challenge lies in the uncertainty within the movie industry and selecting the appropriate machine learning model. We can combine movie features and sentiment analysis from social media using machine learning techniques to achieve movie success prediction. The methods used for predicting based on movie features are Ensemble models (Random Forest + Gradient Boosting). Meanwhile, the methods used for sentiment analysis of trailer comments is LSTM. The evaluation of the models used is based on RMSE and accuracy calculation. The final prediction of success obtains an RMSE of 0,8807 and an accuracy of 91,19%. This represents an improvement from previous research. Further research is recommended to implement the model in the movie industry

Optimization of Sentiment Analysis Classification of ChatGPT on Big Data Twitter in Indonesia using BERT Sinaga, Frans Mikael; Purba, Ronsen; Pipin, Sio Jurnalis; Lestari, Wulan Sri; Winardi, Sunaryo
JURNAL MEDIA INFORMATIKA BUDIDARMA Vol 8, No 3 (2024): Juli 2024
Publisher : Universitas Budi Darma

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.30865/mib.v8i3.7861

This research is grounded in the emergence of ChatGPT technology, supported by prior and similar studies. The urgency of the issue is highlighted by previous research indicating non-convergent classification outcomes in LSTM (Long Short-Term Memory) methods due to suboptimal hyperparameter settings and limitations in understanding text data within Big Data. The presence of ChatGPT technology brings both benefits and potential misuse, such as copyright infringement, unauthorized news extraction, and violations of accountability principles. Understanding public sentiment towards the presence of ChatGPT technology is crucial. The research aims to implement the BERT (Bidirectional Encoder Representations from Transformers) method to achieve accurate and convergent sentiment analysis classification. This study involves data preprocessing stages using Natural Language Processing (NLP) techniques. Text data, already vectorized, is classified using BERT to determine public sentiment (positive, negative, neutral) towards ChatGPT technology, ensuring greater accuracy, convergence, and contextual relevance. Performance testing of the BERT model is conducted using a Confusion Matrix. With parameters set to Max Sequence Length = 128 and Batch Size = 16, the highest classification accuracy achieved is 93.4%.

Implementation of IndoBERT in Sarcasm Detection using Random Forest Towards Sentiment Analysis Sibarani, Sabrina Adela Br; Purba, Ronsen; Limbong, Ricky Paian
Building of Informatics, Technology and Science (BITS) Vol 6 No 4 (2025): March 2025
Publisher : Forum Kerjasama Pendidikan Tinggi

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.47065/bits.v6i4.5801

Sarcasm, a subtle form of irony, often introduces a discrepancy between the literal meaning of words and the intended message, making it a significant challenge for sentiment analysis systems. Misinterpreting sarcasm in social media comments can lead to inaccurate sentiment classification, hindering decision-making processes in areas like customer feedback analysis and social opinion mining. This study addresses this issue by evaluating the effectiveness of sarcasm detection in Indonesian text using a Random Forest Classifier (RFC) integrated with IndoBERT. The research employs 10-fold cross-validation to measure performance. Without IndoBERT, the RFC model achieved average accuracy, precision, recall, and F1-score of 78.83%, 78.83%, 79.01%, and 78.83%, respectively. Incorporating IndoBERT significantly improved performance, with all metrics exceeding 84%. Furthermore, 5-fold cross-validation achieved the highest performance, with all metrics reaching 97.24%. This research contributes to developing more robust natural language processing models tailored to Indonesian linguistic contexts, specifically for sarcasm detection.

Political Comperative Analysis of Indonesian Political Fake News Detection using IndoBERT-Bi-GRU-Attention Models: Evaluating Performance on Narratives and News Headlines Datasets Manurung, Juliana Damayanti; Purba, Ronsen
Building of Informatics, Technology and Science (BITS) Vol 6 No 4 (2025): March 2025
Publisher : Forum Kerjasama Pendidikan Tinggi

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.47065/bits.v6i4.6938

The instant and massive spread of fake news on social media negatively impacts public trust in the media and news agencies. In politics, fake news is often used by politicians to gain support ahead of elections. Detecting fake news in Indonesia poses a significant challenge, especially for communities vulnerable to misinformation. This study aims to develop a new model that combines IndoBERT with Bi-GRU and Attention. Additionally, a comparison is made between the main model and two word embedding models, FastText and GloVe. The tests were conducted on datasets of headlines and news narratives separately. Data was sourced from CNN, Tempo.co, Kompas, and TurnBackHoax.ID. The results show that the IndoBERT-Bi-GRU-Attention model with FastText excelled on the headline dataset with an accuracy of 99.76% and an F1-Score of 99.61%, while the main IndoBERT-Bi-GRU-Attention model excelled on the narrative dataset with an accuracy of 99.08% and an F1-Score of 98.40%. This research demonstrates that IndoBERT can be combined with Bi-GRU, significantly contributing to the development of fake news detection models.

Pengamanan Citra Digital Menggunakan Kriptografi DnaDan Modified LSB Br Sibarani, Sabrina Adela; Munthe, Andreas; Purba, Ronsen; Lubis, Ali Akbar
Jurnal Teknologi Informasi dan Ilmu Komputer Vol 11 No 6: Desember 2024
Publisher : Fakultas Ilmu Komputer, Universitas Brawijaya

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.25126/jtiik.1167666

Enkripsi citra digital menggunakan Kriptografi DNA menggabungkan ilmu komputasi dengan prinsip biologis untuk memberikan keamanan ganda. Proses enkripsi terdiri dari dua lapisan. Lapisan pertama, sistem chaos seperti Arnold's Cat Map (ACM) digunakan untuk mengacak posisi piksel melalui beberapa iterasi, sementara Logistic Map (LM) membangkitkan keystream karena sensitivitasnya yang tinggi. Lapisan kedua melibatkan karakteristik DNA, yang memanfaatkan basa nukleotida (A, T, C, G) untuk mengenkripsi data citra pada tingkat molekuler, menghasilkan tingkat keacakan yang tinggi. Setelah enkripsi, ciphertext disembunyikan dalam citra sampul menggunakan teknik steganografi Modified Least Significant Bit (MLSB), yang mengoptimalkan penyisipan bit di saluran RGB dengan pemilihan piksel acak menggunakan generator modulo. Hasil pengujian menunjukkan kualitas enkripsi yang sangat baik, dengan nilai NPCR ≥ 98%, UACI ≥ 30%, koefisien korelasi ≃ 0, entropi ≃ 8, dan histogram yang datar (flat). Kualitas stego-image optimal dicapai dengan penyisipan satu bit pada saluran RGB, menghasilkan PSNR ≥ 50dB. Ketahanan stego-image terhadap noise salt & pepper bergantung pada ukuran citra sampul, persentase noise, dan jumlah bit sisip yang digunakan. Hasil tersebut menunjukkan bahwa kombinasi Kriptografi DNA, ACM, LM, dan MLSB memberikan keamanan yang tinggi dan sulit ditembus. Abstract Digital image encryption using DNA Cryptography combines computational science with biological principles to provide dual security. The encryption process consists of two layers: first, a chaotic system like Arnold's Cat Map (ACM) is used to shuffle pixel positions through several iterations, while the Logistic Map (LM) generates a keystream due to its high sensitivity. The second layer involves DNA characteristics, utilizing nucleotide bases (A, T, C, G) to encrypt image data at the molecular level, resulting in higher randomness. After encryption, the ciphertext is hidden within a cover image using Modified Least Significant Bit (MLSB) steganography, which optimizes bit insertion in the RGB channels by selecting random pixels using a modulo generator. Experimental results show excellent encryption quality, with NPCR ≥ 98%, UACI ≥ 30%, correlation coefficient close to 0, entropy close to 8, and a flat histogram. Optimal stego-image quality is achieved with a single bit insertion in the RGB channels, resulting in PSNR ≥ 50dB. The resistance of the stego-image to salt & pepper noise depends on the cover image size, noise percentage, and the number of inserted bits. The results indicate that the combination of DNA Cryptography, ACM, LM, and MLSB provides high security and is difficult to breach.

Optimisasi Hyperparameter BiLSTM Menggunakan Bayesian Optimization untuk Prediksi Harga Saham Simamora, Fandi Presly; Purba, Ronsen; Pasha, Muhammad Fermi
Jambura Journal of Mathematics Vol 7, No 1: February 2025
Publisher : Department of Mathematics, Universitas Negeri Gorontalo

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.37905/jjom.v7i1.27166

The accuracy of deep learning models in predicting dynamic and non-linear stock market data highly depends on selecting optimal hyperparameters. However, finding optimal hyperparameters can be costly in terms of the model's objective function, as it requires testing all possible combinations of hyperparameter configurations. This research aims to find the optimal hyperparameter configuration for the BiLSTM model using Bayesian Optimization. The study was conducted using three blue-chip stocks from different sectors, namely BBCA, BYAN, and TLKM, with two scenarios of search iterations. The test results show that Bayesian Optimization was able to find the optimal hyperparameter configuration for the BiLSTM model, with the best MAPE values for each stock: BBCA 1.2092%, BYAN 2.0609%, and TLKM 1.2027%. Compared to previous research on Grid Search-BiLSTM, the use of Bayesian Optimization-BiLSTM resulted in lower MAPE values.

Title

Found 32 Documents
Search

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Title Search

Found 32 Documents Search

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Title

Found 32 Documents
Search