Yulius Denny Prabowo
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Yulius Denny Prabowo has a total of 20 co-authors. Here are the top co-authors by number of shared publications:
- Alfa Yohannis
Faculty of Computer Science and Communication Science, KALBIS Institute, Jakarta, Indonesia
3 shared publications - Martinus Maslim
Universitas Atma Jaya Yogyakarta, Sleman, Indonesia
2 shared publications - Abhishek Sheetal
Faculty of Business, The Hong Kong Polytechnic University , Li Ka Shing Tower, Hung Hom, Kowloon
1 shared publication - Achmad Imam Kistijantoro
Institut Teknologi Bandung, Bandung, Indonesia
1 shared publication - Albert William Wibisono
Universitas Atma Jaya Yogyakarta, Sleman, Indonesia
1 shared publication Alexander Waworuntu
Faculty of Computer Science and Communication Science, KALBIS Institute, Jakarta, Indonesia
1 shared publication
Latest Publications
- Open Access
- 28 October 2026
- Published by Universitas Airlangga in Journal of Information Systems Engineering and Business Intelligence
- Vol. 12 (3) , 580-595
Optimizing Fine-Tuning Memory Usage with QLoRA for Automated Pedagogical Feedback Generation in Bahasa Indonesia
Background: While recent advances in open-source large language models (LLMs) have enabled automated feedback generation for educational assessment, most other research is focused on English and requires high-end Graphics Processing Unit (GPU) resources. However, there is limited empirical evidence of the efficiency of LLMs in Bahasa Indonesia feedback tasks in low-resource settings like Indonesia, given the strict hardware limitations and differences in data sources Objective: This study investigates whether 4-bit quantization combined with Low-Rank Adaptation (LoRA)-based task adaptation can enable open-weight LLMs to generate Indonesian pedagogical feedback under strict 8GB Video Random Access Memory (VRAM) constraints, comparing the performance across various model architectures and supervision data (human-curated vs synthetic). Methods: A quantitative and experimental design using a 2 × 2 factorial structure was used to compare Llama-3-8B-Instruct and Gemma-2-9B-Instruct on 2 datasets of equal size (1,950 human-curated and 1,950 synthetic feedback pairs). For model fine-tuning, we implemented 4-bit Quantized Low-Rank Adaptation (QLoRA) quantization on consumer-grade GPUs with 8GB VRAM Performance was assessed by semantic similarity using Sentence-BERT (SBERT), similarity of the model with BLEU-4, ROUGE-L, peak VRAM usage, and training time. Results: QLoRA successfully enabled fine-tuning within 8GB VRAM, with peak memory usage between 7.0 and 7.8GB. Gemma-2-9B-Instruct achieved the highest semantic alignment (SBERT similarity = 0.77 on synthetic and 0.74 on human data), while Llama-3-8B-Instruct showed competitive performance (0.76 and 0.71, respectively). Human-curated datasets produced higher lexical alignment (ROUGE-1 up to 44.8) than synthetic datasets, though synthetic supervision remained semantically consistent. Conclusion: QLoRA enables efficient fine-tuning of large language models on consumer-grade hardware, making Indonesian feedback generation feasible in low-resource settings. While 4-bit quantization reduces memory requirements, LoRA enables task-specific pedagogical adaptation. Instruction-tuned models consistently outperform language-adapted base models. Future work should include human evaluation and broader educational contexts to improve generalizability. Keywords: Automated Feedback Generation, Large Language Models, QLoRA Fine-Tuning, Bahasa Indonesia, Educational Assessment
- 28 October 2026
- Published by Universitas Airlangga in Journal of Information Systems Engineering and Business Intelligence
- Vol. 12 (3) , 580-595
Open AccessOptimizing Fine-Tuning Memory Usage with QLoRA for Automated Pedagogical Feedback Generation in Bahasa Indonesia
Background: While recent advances in open-source large language models (LLMs) have enabled automated feedback generation for educational assessment, most other research is focused on English and requires high-end Graphics Processing Unit (GPU) resources. However, there is limited empirical evidence of the efficiency of LLMs in Bahasa Indonesia feedback tasks in low-resource settings like Indonesia, given the strict hardware limitations and differences in data sources Objective: This study investigates whether 4-bit quantization combined with Low-Rank Adaptation (LoRA)-based task adaptation can enable open-weight LLMs to generate Indonesian pedagogical feedback under strict 8GB Video Random Access Memory (VRAM) constraints, comparing the performance across various model architectures and supervision data (human-curated vs synthetic). Methods: A quantitative and experimental design using a 2 × 2 factorial structure was used to compare Llama-3-8B-Instruct and Gemma-2-9B-Instruct on 2 datasets of equal size (1,950 human-curated and 1,950 synthetic feedback pairs). For model fine-tuning, we implemented 4-bit Quantized Low-Rank Adaptation (QLoRA) quantization on consumer-grade GPUs with 8GB VRAM Performance was assessed by semantic similarity using Sentence-BERT (SBERT), similarity of the model with BLEU-4, ROUGE-L, peak VRAM usage, and training time. Results: QLoRA successfully enabled fine-tuning within 8GB VRAM, with peak memory usage between 7.0 and 7.8GB. Gemma-2-9B-Instruct achieved the highest semantic alignment (SBERT similarity = 0.77 on synthetic and 0.74 on human data), while Llama-3-8B-Instruct showed competitive performance (0.76 and 0.71, respectively). Human-curated datasets produced higher lexical alignment (ROUGE-1 up to 44.8) than synthetic datasets, though synthetic supervision remained semantically consistent. Conclusion: QLoRA enables efficient fine-tuning of large language models on consumer-grade hardware, making Indonesian feedback generation feasible in low-resource settings. While 4-bit quantization reduces memory requirements, LoRA enables task-specific pedagogical adaptation. Instruction-tuned models consistently outperform language-adapted base models. Future work should include human evaluation and broader educational contexts to improve generalizability. Keywords: Automated Feedback Generation, Large Language Models, QLoRA Fine-Tuning, Bahasa Indonesia, Educational Assessment
- Open Access
- 28 October 2025
- Published by Universitas Airlangga in Journal of Information Systems Engineering and Business Intelligence
- Vol. 11 (3) , 500-513
Diabetic Retinopathy Fundus Image Classification Using Self-Organizing Maps
Background: Diabetic retinopathy (DR) is a condition that impairs the blood vessels in the retina, resulting in vision loss ranging from temporary to permanent blindness. It commonly affects individuals diagnosed with diabetes mellitus (DM). Fundoscopy is a technique used to identify DR by examining the fundus of the eye during an eye examination. This process is time-consuming and can be expensive. Objective: This study aimed to examine the identification of DR using digital image processing methods. Methods: The self-organizing map (SOM) artificial neural network was employed. Diabetic retinopathy will be categorized according to its severity, including normal, mild, moderate, or severe. This classification considers the quantity of exudates and microaneurysms and the blood vessel structure in the fundus image. The dataset used in this investigation comprised 1000 fundus images acquired from the MESSIDOR ophthalmology database. Results: The findings indicate that the SOM approach achieves a training accuracy of 72% and a testing accuracy of 62%. Conclusion: The DR severity classification system can effectively extract DR-related features by segmenting exudates, blood vessels, and microaneurysms from funduscopic images during training, testing, and evaluation. Keywords: Diabetic Retinopathy, Self-Organizing Map, Fundus Image Classification, Digital Image Processing
- 28 October 2025
- Published by Universitas Airlangga in Journal of Information Systems Engineering and Business Intelligence
- Vol. 11 (3) , 500-513
Open AccessDiabetic Retinopathy Fundus Image Classification Using Self-Organizing Maps
Background: Diabetic retinopathy (DR) is a condition that impairs the blood vessels in the retina, resulting in vision loss ranging from temporary to permanent blindness. It commonly affects individuals diagnosed with diabetes mellitus (DM). Fundoscopy is a technique used to identify DR by examining the fundus of the eye during an eye examination. This process is time-consuming and can be expensive. Objective: This study aimed to examine the identification of DR using digital image processing methods. Methods: The self-organizing map (SOM) artificial neural network was employed. Diabetic retinopathy will be categorized according to its severity, including normal, mild, moderate, or severe. This classification considers the quantity of exudates and microaneurysms and the blood vessel structure in the fundus image. The dataset used in this investigation comprised 1000 fundus images acquired from the MESSIDOR ophthalmology database. Results: The findings indicate that the SOM approach achieves a training accuracy of 72% and a testing accuracy of 62%. Conclusion: The DR severity classification system can effectively extract DR-related features by segmenting exudates, blood vessels, and microaneurysms from funduscopic images during training, testing, and evaluation. Keywords: Diabetic Retinopathy, Self-Organizing Map, Fundus Image Classification, Digital Image Processing
- 26 April 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Computer
- Vol. 56 (5) , 74-83
Functional Analogies Increase Trust in Black-Box AI Systems Among Lay Consumers: The Case of GeNose C-19
We present a case study of a black-box artificial intelligence-based COVID-19 detection product, GeNose C-19, developed by the Indonesian government. We find that explaining how GeNose works using functional analogies increases both Indonesian and American lay consumers’ trust in GeNose.
- 26 April 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Computer
- Vol. 56 (5) , 74-83
Functional Analogies Increase Trust in Black-Box AI Systems Among Lay Consumers: The Case of GeNose C-19
We present a case study of a black-box artificial intelligence-based COVID-19 detection product, GeNose C-19, developed by the Indonesian government. We find that explaining how GeNose works using functional analogies increases both Indonesian and American lay consumers’ trust in GeNose.
- Open Access
- 26 April 2019
- Published by Universitas Atma Jaya Yogyakarta in Jurnal Buana Informatika
- Vol. 10 (1) , 29-40
Pembentukan Vector Space Model Bahasa Indonesia Menggunakan Metode Word to Vector
Extracting information from a large amount of structured data requires expensive computing. The Vector Space Model method works by mapping words in continuous vector space where semantically similar words are mapped in adjacent vector spaces. The Vector Space Model model assumes words that appear in the same context, having the same semantic meaning. In the implementation, there are two different approaches: counting methods (eg: Latent Semantic Analysis) and predictive methods (eg Neural Probabilistic Language Model). This study aims to apply Word2Vec method using the Continuous Bag of Words approach in Indonesian language. Research data was obtained by crawling on several online news portals. The expected result of the research is the Indonesian words vector mapping based on the data used.Keywords: vector space model, word to vector, Indonesian vector space model.Abstrak. Ekstraksi informasi dari sekumpulan data terstruktur dalam jumlah yang besar membutuhkan komputasi yang mahal. Metode Vector Space Model bekerja dengan cara memetakan kata-kata dalam ruang vektor kontinu dimana kata-kata yang serupa secara semantis dipetakan dalam ruang vektor yang berdekatan. Metode Vector Space Model mengasumsikan kata-kata yang muncul pada konteks yang sama, memiliki makna semantik yang sama. Dalam penerapannya ada dua pendekatan yang berbeda yaitu: metode yang berbasis hitungan (misal: Latent Semantic Analysis) dan metode prediktif (misalnya Neural Probabilistic Language Model). Penelitian ini bertujuan untuk menerapkan metode Word2Vec menggunakan pendekatan Continuous Bag Of Words model dalam Bahasa Indonesia. Data penelitian yang digunakan didapatkan dengan cara crawling pada berberapa portal berita online. Hasil penelitian yang diharapkan adalah pemetaan vektor kata Bahasa Indonesia berdasarkan data yang digunakan.Kata Kunci: vector space model, word to vector, vektor kata bahasa Indonesia.
- 26 April 2019
- Published by Universitas Atma Jaya Yogyakarta in Jurnal Buana Informatika
- Vol. 10 (1) , 29-40
Open AccessPembentukan Vector Space Model Bahasa Indonesia Menggunakan Metode Word to Vector
Extracting information from a large amount of structured data requires expensive computing. The Vector Space Model method works by mapping words in continuous vector space where semantically similar words are mapped in adjacent vector spaces. The Vector Space Model model assumes words that appear in the same context, having the same semantic meaning. In the implementation, there are two different approaches: counting methods (eg: Latent Semantic Analysis) and predictive methods (eg Neural Probabilistic Language Model). This study aims to apply Word2Vec method using the Continuous Bag of Words approach in Indonesian language. Research data was obtained by crawling on several online news portals. The expected result of the research is the Indonesian words vector mapping based on the data used.Keywords: vector space model, word to vector, Indonesian vector space model.Abstrak. Ekstraksi informasi dari sekumpulan data terstruktur dalam jumlah yang besar membutuhkan komputasi yang mahal. Metode Vector Space Model bekerja dengan cara memetakan kata-kata dalam ruang vektor kontinu dimana kata-kata yang serupa secara semantis dipetakan dalam ruang vektor yang berdekatan. Metode Vector Space Model mengasumsikan kata-kata yang muncul pada konteks yang sama, memiliki makna semantik yang sama. Dalam penerapannya ada dua pendekatan yang berbeda yaitu: metode yang berbasis hitungan (misal: Latent Semantic Analysis) dan metode prediktif (misalnya Neural Probabilistic Language Model). Penelitian ini bertujuan untuk menerapkan metode Word2Vec menggunakan pendekatan Continuous Bag Of Words model dalam Bahasa Indonesia. Data penelitian yang digunakan didapatkan dengan cara crawling pada berberapa portal berita online. Hasil penelitian yang diharapkan adalah pemetaan vektor kata Bahasa Indonesia berdasarkan data yang digunakan.Kata Kunci: vector space model, word to vector, vektor kata bahasa Indonesia.
- 1 September 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 51-56
Lstm And Simple Rnn Comparison In The Problem Of Sequence To Sequence On Conversation Data Using Bahasa Indonesia
This study aims to implement and compare the Long Short Term Memory (LSTM) and Simple Recurrent Neural Networks (RNN) algorithm in the case of chatbot using Bahasa Indonesia data. The chatbot model used is a cahatbot model across business/service fields. The training data used in this research are the data on customer service talks with its customers in several business fields or services. To compare the models generated from the LSTM algorithm and Simple RNN algorithm, two tests were carried out, the first test is testing the chat output manually which was done directly by humans and the second test are comparing the LSTM algorithm and Simple RNN algorithms using the same training data and test data. From the experimental results, it was found that the chat output generated by the LSTM algorithm relatively can answer most of the tests correctly rather than Simple RNN algorithm. From the experiment, it was also found that the learning process in the LSTM algorithm takes longer than the learning process on the Simple RNN algorithm.
- 1 September 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 51-56
Lstm And Simple Rnn Comparison In The Problem Of Sequence To Sequence On Conversation Data Using Bahasa Indonesia
This study aims to implement and compare the Long Short Term Memory (LSTM) and Simple Recurrent Neural Networks (RNN) algorithm in the case of chatbot using Bahasa Indonesia data. The chatbot model used is a cahatbot model across business/service fields. The training data used in this research are the data on customer service talks with its customers in several business fields or services. To compare the models generated from the LSTM algorithm and Simple RNN algorithm, two tests were carried out, the first test is testing the chat output manually which was done directly by humans and the second test are comparing the LSTM algorithm and Simple RNN algorithms using the same training data and test data. From the experimental results, it was found that the chat output generated by the LSTM algorithm relatively can answer most of the tests correctly rather than Simple RNN algorithm. From the experiment, it was also found that the learning process in the LSTM algorithm takes longer than the learning process on the Simple RNN algorithm.
- 1 March 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 773-777
An initial research on Halstead's technique for pattern similarity relationship study
Halstead Complexity Measures is a software metrics introducted by Maurice Howard Halstead in 1977. This metrics used to identify measurable properties of the code and relation between them. Several measures can be calculated with this method are program vocabulary, program length, volume, difficulty and effort. In learning process, the final result obtained by students should be linear with the results obtained each time the students doing their exercises. This study aims to see the pattern similarity relationship between coding that students do on each exercise in term of difficulty, programming effort and programming time with the final value that they obtained. The experiment was done by giving the students three sets of exercise questions which were derived from the final problem, then each set of exercise code done by each students measured using halstead method and compare with the final grade obtained by each student.
- 1 March 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 773-777
An initial research on Halstead's technique for pattern similarity relationship study
Halstead Complexity Measures is a software metrics introducted by Maurice Howard Halstead in 1977. This metrics used to identify measurable properties of the code and relation between them. Several measures can be calculated with this method are program vocabulary, program length, volume, difficulty and effort. In learning process, the final result obtained by students should be linear with the results obtained each time the students doing their exercises. This study aims to see the pattern similarity relationship between coding that students do on each exercise in term of difficulty, programming effort and programming time with the final value that they obtained. The experiment was done by giving the students three sets of exercise questions which were derived from the final problem, then each set of exercise code done by each students measured using halstead method and compare with the final grade obtained by each student.
- 1 October 2015
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 56-61
Visualization of user-learning game interaction unveiling learner's learning patterns
Many existing games collect various data from their users and some of the data are in the form of event logs. Event logs of games are potential resources to obtain empirical data since they capture the real world of how games are used by players. Using the event logs, we can discover the behavioral models of how players interact with games. Specifically for learning games, their event logs can be used to discover the learning patterns of learners. From the use of our sorting algorithm learning artifact, an event log is collected and used to construct a player's activity graph during his or her interaction with the game. To identify the learner's learning patterns, the graph is visualized using Harel-Koren Fast Multiscale and Horizontal Sine layouts. Some visual spatial attributes of nodes and edges that construct the graph are used to convey several information about the course of events resulted from the interaction between the learner and the artifact. Based on the visualization, this research has successfully identified ten interaction patterns, in which most of them exhibit that the learner whose activities are logged tends to perform learning based on the experiential learning rather than the didactic learning. The learner prefers exerting “trial and error” approach to build an association rule in order to predict the next correct steps in executing the sorting algorithm. The identified interaction patterns will be studied further into formalization and the construction of algorithms to identify the learning patterns in the real-world.
- 1 October 2015
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 56-61
Visualization of user-learning game interaction unveiling learner's learning patterns
Many existing games collect various data from their users and some of the data are in the form of event logs. Event logs of games are potential resources to obtain empirical data since they capture the real world of how games are used by players. Using the event logs, we can discover the behavioral models of how players interact with games. Specifically for learning games, their event logs can be used to discover the learning patterns of learners. From the use of our sorting algorithm learning artifact, an event log is collected and used to construct a player's activity graph during his or her interaction with the game. To identify the learner's learning patterns, the graph is visualized using Harel-Koren Fast Multiscale and Horizontal Sine layouts. Some visual spatial attributes of nodes and edges that construct the graph are used to convey several information about the course of events resulted from the interaction between the learner and the artifact. Based on the visualization, this research has successfully identified ten interaction patterns, in which most of them exhibit that the learner whose activities are logged tends to perform learning based on the experiential learning rather than the didactic learning. The learner prefers exerting “trial and error” approach to build an association rule in order to predict the next correct steps in executing the sorting algorithm. The identified interaction patterns will be studied further into formalization and the construction of algorithms to identify the learning patterns in the real-world.
- 1 September 2015
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-8
Sort Attack: Visualization and Gamification of Sorting Algorithm Learning
Algorithms are commonly perceived as a difficult subject, which is quite an irony as they have a fundamental role in computer science. Failure to master this subject will inhibit students'' capabilities as they advance to higher levels. Algorithm visualization, as an effort to overcome the problem, has been growing towards gameful visualization recently that is presumed to be able to engage learners longer and more intensely. However, integrating algorithm visualization, game elements, and instructional design is not a trivial task as it requires a careful design. Hence, a conceptual model of how algorithm learning instructions, algorithm visualization, and gamification improve learning outcomes was developed. While instructional design concerns with developing the best strategy for learning, algorithm visualization functions as a cognitive support provider and gamification works by delivering engagement. Both cognitive support and engagement moderate the learning instructions that concern with enhancing learning outcomes. Principles and elements originating from the three domains have to be taken into consideration during the designing process to produce an artifact that can deliver the essential functions of each domain. A working artifact was then constructed, as the instantiation of the model, to validate whether the idea of integrating algorithm visualization and gamification into algorithm learning instructions is capable enough to improve learning outcomes. Based on our respondents'' learning outcomes, it was found the artifact can significantly improve the procedural knowledge of learners that are indicated by their increased capability in solving sorting algorithm problems.
- 1 September 2015
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-8
Sort Attack: Visualization and Gamification of Sorting Algorithm Learning
Algorithms are commonly perceived as a difficult subject, which is quite an irony as they have a fundamental role in computer science. Failure to master this subject will inhibit students'' capabilities as they advance to higher levels. Algorithm visualization, as an effort to overcome the problem, has been growing towards gameful visualization recently that is presumed to be able to engage learners longer and more intensely. However, integrating algorithm visualization, game elements, and instructional design is not a trivial task as it requires a careful design. Hence, a conceptual model of how algorithm learning instructions, algorithm visualization, and gamification improve learning outcomes was developed. While instructional design concerns with developing the best strategy for learning, algorithm visualization functions as a cognitive support provider and gamification works by delivering engagement. Both cognitive support and engagement moderate the learning instructions that concern with enhancing learning outcomes. Principles and elements originating from the three domains have to be taken into consideration during the designing process to produce an artifact that can deliver the essential functions of each domain. A working artifact was then constructed, as the instantiation of the model, to validate whether the idea of integrating algorithm visualization and gamification into algorithm learning instructions is capable enough to improve learning outcomes. Based on our respondents'' learning outcomes, it was found the artifact can significantly improve the procedural knowledge of learners that are indicated by their increased capability in solving sorting algorithm problems.
- 1 November 2014
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 284-289
Defining gamification: From lexical meaning and process viewpoint towards a gameful reality
In a state where games and their elements have been extensively used not only for pleasure but also for other purposes, gamification still has some pros and cons about its definition, which might influence people's decisions on their game-related strategies to improve their performance. This work tries to define gamification by using lexical meaning approach as the starting point and viewing it from a process viewpoint. Lexical meaning approach interprets gamification as a process or a product of the process. From this perspective, gamification can be viewed as a process that adds certain characteristics to an object that makes the object different from its previous condition and feasible to be formalized. Furthermore, the resulting definition is tested by comparing it to other existing gamification definitions and the understanding that constructs the definition is used as the foundation to explain the differences between gamification and serious games. This paper then defines gamification as a process that integrates game elements into gameless objects in order to have gameful characteristics. There will be a situation where gamification will produce a state of gameful reality: a situation in the real world where people can feel the significant presence of gamefulness in their daily life.
- 1 November 2014
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 284-289
Defining gamification: From lexical meaning and process viewpoint towards a gameful reality
In a state where games and their elements have been extensively used not only for pleasure but also for other purposes, gamification still has some pros and cons about its definition, which might influence people's decisions on their game-related strategies to improve their performance. This work tries to define gamification by using lexical meaning approach as the starting point and viewing it from a process viewpoint. Lexical meaning approach interprets gamification as a process or a product of the process. From this perspective, gamification can be viewed as a process that adds certain characteristics to an object that makes the object different from its previous condition and feasible to be formalized. Furthermore, the resulting definition is tested by comparing it to other existing gamification definitions and the understanding that constructs the definition is used as the foundation to explain the differences between gamification and serious games. This paper then defines gamification as a process that integrates game elements into gameless objects in order to have gameful characteristics. There will be a situation where gamification will produce a state of gameful reality: a situation in the real world where people can feel the significant presence of gamefulness in their daily life.