Abstract:Abstract: This study discusses the optimization of RAG for a FAQ system in the field of information technology product security certification at BSSN. Although LLM generate reliable responses, they often lack up-to-date…
and domain-specific knowledge, which can be addressed through the RAG approach. This research aims to optimize a domain-specific RAG system by improving embedding performance, enhancing prompt robustness, and increasing retrieval accuracy. The research methods consist of three stages. The first stage involves fine-tuning the bge-m3 embedding model and evaluating its performance using MRR, Recall, and AUC. The second stage applies prompt engineering techniques, namely the SRSM and Autodefense, to mitigate direct-injection and escape-character prompt injection attacks. The third stage evaluates the proposed RAG system using Precision, Recall, and F1-Score metrics against four baseline models. The results of research show that the fine-tuned embedding model achieves higher performance than the original model, with MRR@1 and Recall@1 values of 0.80 and an AUC@100 of 0.7023. In addition, the proposed prompt engineering techniques demonstrate robustness against prompt injection attacks, while the overall RAG system attains a perfect Precision, Recall, and F1-Score of 1.00. In conclusion, the proposed approach effectively enhances retrieval accuracy, embedding quality, and system security, resulting in a more reliable RAG-based FAQ system for information technology product security certification.
Keywords: embedding fine-tuning; large language model; prompt engineering; prompt injection mitigation; retrieval-augmented generation
Abstrak: Studi ini membahas optimasi RAG untuk sistem FAQ di bidang sertifikasi keamanan produk teknologi informasi di BSSN. Meskipun LLM menghasilkan respons yang andal, mereka seringkali kurang memiliki pengetahuan terkini dan spesifik domain, yang dapat diatasi melalui pendekatan RAG. Penelitian ini bertujuan untuk mengoptimalkan sistem RAG spesifik domain dengan meningkatkan kinerja embedding, meningkatkan ketahanan prompt dan meningkatkan akurasi pengambilan. Metode penelitian terdiri dari tiga tahap. Tahap pertama melibatkan fine-tuning model embedding bge-m3 dan mengevaluasi kinerjanya menggunakan Mean Reciprocal Rank (MRR), Recall, dan AUC. Tahap kedua menerapkan teknik rekayasa prompt, yaitu Self- SRSM dan Autodefense, untuk mengurangi serangan direct-injection dan escape-character prompt injection. Tahap ketiga mengevaluasi sistem RAG yang diusulkan menggunakan metrik Presisi, Recall, dan F1-Score terhadap empat model dasar. Hasil penelitian menunjukkan bahwa model embedding yang disempurnakan mencapai kinerja yang lebih tinggi daripada model asli, dengan nilai MRR@1 dan Recall@1 sebesar 0,80 dan AUC@100 sebesar 0,7023. Selain itu, teknik rekayasa prompt yang diusulkan menunjukkan ketahanan terhadap serangan injeksi prompt, sementara sistem RAG secara keseluruhan mencapai Presisi, Recall, dan F1-Score sempurna sebesar 1,00. Kesimpulannya, pendekatan yang diusulkan secara efektif meningkatkan akurasi pengambilan, kualitas embedding dan keamanan sistem, menghasilkan sistem FAQ berbasis RAG yang lebih andal untuk sertifikasi keamanan produk teknologi informasi.
Kata kunci: penyempurnaan embedding; model bahasa besar; rekayasa prompt; mitigasi injeksi prompt; retrieval-augmented generation
Abstract:Abstract: Obesity is an escalating global health concern, with unhealthy lifestyle patterns contributing significantly to its development. This study aims to evaluate and compare three clustering techniques for categorizing…
ing lifestyle patterns and obesity-related factors: K-Means, Agglomerative Clustering, and Gaussian Mixture Model (GMM). The data used in this study is sourced from the Food Nutrition dataset, which includes variables such as dietary habits, physical activity, and socio-economic status. The three clustering methods were assessed using evaluation metrics such as Silhouette Score, Davies-Bouldin Index (DBI), and Calinski-Harabasz Index (CHI). The findings revealed that K-Means exhibited the best performance in terms of cluster separation with a Silhouette Score of 0.5559, while GMM showed better flexibility in handling more complex data. Although Agglomerative Clustering produced acceptable results, it had a higher overlap between clusters compared to the other methods. This study offers valuable insights into selecting the most appropriate clustering technique based on the data characteristics.
Keywords: agglomerative; clustering; GMM; k-means; lifestyle patterns; obesity
Abstrak: Obesitas menjadi masalah kesehatan yang semakin meningkat di seluruh dunia, dengan pola hidup yang tidak sehat berperan besar dalam perkembangannya. Penelitian ini bertujuan untuk membandingkan tiga metode clustering dalam mengelompokkan pola gaya hidup dan faktor yang memengaruhi obesitas, yaitu K-Means, Agglomerative Clustering, dan Gaussian Mixture Model (GMM). Data yang digunakan diperoleh dari dataset Food Nutrition yang mencakup informasi terkait pola makan, aktivitas fisik, serta faktor sosial-ekonomi. Ketiga metode tersebut diuji dengan menggunakan beberapa metrik evaluasi, seperti Silhouette Score, Davies-Bouldin Index (DBI), dan Calinski-Harabasz Index (CHI). Hasil penelitian menunjukkan bahwa K-Means memiliki kinerja terbaik dalam hal pemisahan klaster, dengan nilai Silhouette Score sebesar 0.5559, sementara GMM lebih fleksibel dalam menangani data yang lebih kompleks. Meskipun Agglomerative Clustering memberikan hasil yang dapat diterima, tumpang tindih antar klaster lebih besar dibandingkan dengan kedua metode lainnya. Penelitian ini memberikan pemahaman yang lebih baik mengenai pemilihan metode clustering yang tepat berdasarkan karakteristik data yang digunakan.
Kata kunci: agglomerative; clustering; GMM; k-means; obesitas; pola gaya hidup
Abstract:Abstract: The policy to increase the Rural and Urban Land and Building Tax (PBB-P2) in Indonesia often elicits mixed reactions from the public. Some support it because they believe it can strengthen regional fiscal capacity,…
ity, while others reject it because they are concerned that it will increase the economic burden on the community. Understanding public sentiment towards this policy is important for evaluating the effectiveness of the policy and formulating appropriate communication strategies. This study aims to analyze public sentiment towards the PBB-P2 increase policy using data uploaded on Platform X (Twitter). The data were collected through crawling with the keyword “building tax,” then processed through several preprocessing stages before classifying tweets into positive and negative sentiments. Two models were used: Support Vector Machine (SVM) and Bidirectional Long Short-Term Memory (BiLSTM). Results show that SVM outperformed BiLSTM, achieving training accuracy of 99.4% and testing accuracy of 85.9%, with accuracy 0.8595, precision 0.8536, recall 0.8595, and F1-score 0.8449. Meanwhile, BiLSTM achieved training accuracy of 86.9% and testing accuracy of 82.9%, with accuracy 0.8294, precision 0.8150, recall 0.8294, and F1-score 0.8080. These findings suggest SVM is more effective in classifying public sentiment and can support better evaluation of regional tax policies.
Keywords: sentiment analysis; PBB-P2; BiLSTM; SVM; X platform
Abstrak: Kebijakan kenaikan tarif Pajak Bumi dan Bangunan Perdesaan dan Perkotaan (PBB-P2) di In-donesia sering memunculkan beragam reaksi dari masyarakat. Sebagian mendukung karena dianggap dapat memperkuat kapasitas fiskal daerah, sementara lainnya menolak karena kha-watir menambah beban ekonomi masyarakat. Pemahaman terhadap sentimen publik atas ke-bijakan tersebut penting untuk mengevaluasi efektivitas kebijakan dan merumuskan strategi komunikasi yang tepat. Penelitian ini bertujuan menganalisis sentimen masyarakat terhadap kebijakan kenaikan PBB-P2 menggunakan data unggahan di Platform X (Twitter). Data dik-umpulkan melalui proses crawling dengan kata kunci “pajak bangunan” kemudian diproses melalui beberapa tahap preprocessing sebelum diklasifikasikan menjadi sentimen positif dan negatif. Dua model digunakan dalam penelitian ini, yaitu Support Vector Machine (SVM) dan Bidirectional Long Short-Term Memory (BiLSTM). Hasil penelitian menunjukkan bahwa SVM memiliki kinerja lebih baik dibandingkan BiLSTM, dengan akurasi pelatihan 99,4% dan akurasi pengujian 85,9%. Nilai akurasi 0,8595, precision 0,8536, recall 0,8595, dan F1-score 0,8449. Sementara itu, BiLSTM memperoleh akurasi pelatihan 86,9% dan akurasi pengujian 82,9%, dengan akurasi 0,8294, precision 0,8150; recall 0,8294; dan F1-score 0,8080. Temuan ini menunjukkan bahwa SVM lebih efektif dalam mengklasifikasikan sentimen publik serta dapat mendukung evaluasi kebijakan pajak daerah dengan lebih baik.
Kata kunci: analisis sentimen; PBB-P2; BiLSTM; SVM; platform X
Abstract:Abstract: The rapid growth of the digital music industry requires accurate music genre classification systems to enhance user experience in streaming services. This study compares a domain-specific Long Short-Term Memory…
(LSTM) network with three Large Language Models (LLMs)—HuBERT, WavLM, and WAV2Vec 2.0—for Music Genre Classification (MGC). The LSTM model was trained using Mel-spectrograms transformed from the GTZAN dataset, while the LLMs were fine-tuned using a smaller set of raw audio samples due to computational constraints. All models were tested on datasets with identical genre labels to ensure a fair evaluation. Results show that the LSTM model achieved the highest accuracy of 97.10%, outperforming HuBERT (86.00%), WavLM (83.00%), and WAV2Vec 2.0 (80.00%). The LSTM demonstrated superior generalization and stability without overfitting, while the LLMs struggled to differentiate between genres with similar acoustic characteristics. These findings indicate that general-purpose pre-trained models, although powerful, are less effective in music-specific tasks due to domain mismatch. Therefore, incorporating music-specific features and architectures remains essential for achieving higher accuracy and reliability in automatic genre classification systems.
Keywords: audio large language models; comparative deep learning; music genre classification.
Abstrak: Pertumbuhan industri musik digital yang pesat menuntut sistem klasifikasi genre musik yang akurat untuk meningkatkan pengalaman pengguna dalam layanan streaming. Penelitian ini dilatarbelakangi oleh perkembangan pesat model pembelajaran mendalam, khususnya jaringan LSTM dan model bahasa berskala besar LLM seperti HuBERT, WavLM, dan WAV2Vec 2.0, yang telah menunjukkan kemampuan representasi audio yang kuat. Tujuan penelitian ini ini membandingkan jaringan Long Short-Term Memory (LSTM) khusus domain dengan tiga model Large Language Models (LLM)—HuBERT, WavLM, dan WAV2Vec 2.0—untuk tugas Klasifikasi Genre Musik (MGC). Metode penelitian melibatkan pelatihan LSTM menggunakan data Mel-spectrogram hasil transformasi dari dataset GTZAN, sementara LLM disesuaikan (fine-tuning) menggunakan data audio mentah dalam jumlah lebih kecil karena keterbatasan komputasi. Seluruh model diuji pada dataset dengan label genre yang sama untuk memastikan evaluasi yang adil. Hasil penelitian menunjukkan bahwa model LSTM mencapai akurasi tertinggi sebesar 97,10%, sedangkan model HuBERT, WavLM, dan WAV2Vec 2.0 masing-masing memperoleh 86,00%, 83,00%, dan 80,00%. Model LSTM menunjukkan kemampuan generalisasi yang lebih baik tanpa overfitting, sedangkan model LLM cenderung kesulitan membedakan genre dengan karakteristik akustik yang mirip. Kesimpulan penelitian ini adalah ketidaksesuaian domain secara signifikan membatasi performa model umum saat diterapkan pada tugas berbasis musik. Oleh karena itu, penggunaan fitur dan arsitektur khusus musik sangat penting dalam membangun sistem klasifikasi genre yang lebih akurat.
Kata kunci: klasifikasi genre musik; model bahasa besar; perbandingan pembelajaran mendalam.
Abstract:Abstract: The rapid growth of the cosmetics industry on e-commerce platforms has intensified competition, creating a critical need for effective, data-driven marketing strategies. This study aims to conduct a comparative…
analysis of machine learning algorithms to predict the sales categories (High, Medium, Low) of cosmetic products on the Tokopedia marketplace. Four classification models; Random Forest, XGBoost, Logistic Regression, and Naive Bayes were trained and evaluated on data collected via web scraping. The methodology incorporates the Synthetic Minority Over-sampling Technique (SMOTE) to address significant class imbalance and GridSearchCV for hyperparameter optimization to ensure a fair and robust comparison. The experimental results conclusively show that the Random Forest model achieved the best performance, yielding the highest F1-Score Macro Average of 0.75 and an accuracy of 85.3%. The superior model was subsequently implemented in a simple recommendation system to simulate optimal discount strategies, demonstrating its practical utility in providing actionable insights for business decisions.
Keywords: classification; comparative analysis; machine learning; sales prediction; SMOTE
Abstrak: Pertumbuhan pesat industri kosmetik pada platform e-commerce telah membuat persaingan ketat, sehingga menciptakan kebutuhan krusial akan strategi pemasaran yang efektif dan berbasis data. Penelitian ini bertujuan untuk melakukan analisis komparatif terhadap algoritma machine learning untuk memprediksi kategori penjualan (Tinggi, Sedang, Rendah) produk kosmetik di marketplace Tokopedia. Empat model klasifikasi, yaitu Random Forest, XGBoost, Regresi Logistik, dan Naive Bayes, dilatih dan dievaluasi menggunakan data yang dikumpulkan melalui web scraping. Metodologi penelitian ini menerapkan Synthetic Minority Over-sampling Technique (SMOTE) untuk mengatasi ketidakseimbangan kelas yang signifikan dan GridSearchCV untuk optimisasi hyperparameter guna memastikan perbandingan yang adil. Hasil eksperimen menunjukkan bahwa model Random Forest mencapai performa terbaik, dengan menghasilkan F1-Score Macro Average tertinggi sebesar 0,75 dan akurasi 85,3%. Model unggul ini kemudian diimplementasikan dalam sebuah sistem rekomendasi sederhana untuk menyimulasikan strategi diskon yang optimal, yang menunjukkan kegunaan praktisnya dalam memberikan wawasan yang dapat ditindaklanjuti untuk pengambilan keputusan bisnis.
Kata kunci: analisis komparatif; klasifikasi; machine learning; prediksi penjualan; SMOTE
Abstract:Abstract: MBG is a strategic program of the Prabowo-Gibran administration. This program has become a widely discussed issue in the public. To better understand public perception of this program, sentiment analysis is necessary.…
essary. This study aims to compare the performance of algorithms machine learning SVM, RF, And BERT with preprocessing data analyzing public sentiment of the MBG program in media X. The total dataset for this study was 39,858 out of 42,465 successfully crawled tweets. The research methods included data collection, preprocessing data (cleaning, case folding, word normalization, stopword removal and stemming), feature extraction, model training (fine-tuning), handling class imbalance with SMOTE, and evaluation using accuracy, precision, recall, and f1-score. The research results show that without SMOTE, the best performing models are BERT with 89% accuracy, SVM 87%, and RF 78.4%. After SMOTE, the best algorithms were SVM with 92.94%, BERT with 88.3%, and RF with 86.59%. The results confirmed that SVM is the best algorithm if at leastclass imbalance. BERT is the best algorithm before and after SMOTE, because BERT is more effective in capturing the nuances of language on social media, so BERT is the most recommended in MBG sentiment analysis.
Keywords: sentiment analysis; machine learning; SVM, RF, and BERT
Abstrak: MBG merupakan program strategis pemerintahan Prabowo - Gibran. Program ini menjadi isu yang banyak diperbincangkan publik. Untuk mengetahui lebih dalam persepsi masyrakat tentang program ini, perlu dilakukan analisis sentiment. Penelitian ini bertujuan membandingkan kinerja algoritma machine learning SVM, RF, dan BERT dengan preprocessing data menganalisis sentiment public program MBG di media X. Total dataset penelitian ini adalah 39.858 dari 42.465 tweet yang berhasil di crawling. Metode penelitian mencakup pengumpulan data, preprocessing data (cleaning, case folding, normalisasi kata, stopword removal dan stemming), ekstraksi fitur, pelatihan model (fine-tuning), penanganan class imbalance dengan SMOTE, dan evaluasi menggunakan akurasi, presisi, recall, dan f1-score. Hasil peneltian menunjukkan, tanpa SMOTE model dengan kinerja terbaik adalah BERT dengan akurasi 89%, SVM 87%, dan RF 78,4%. Setelah SMOTE algoritma terbaik adalah SVM 92,94%, BERT 88,3% dan RF 86,59%. Hasil penelitian menegaskan bahwa SVM adalah algoritma terbaik jika minimal class imbalance. BERT adalah algoritma terbaik sebelum dan sesudah SMOTE, karena BERT lebih efektif dalam menangkap nuansa bahasa pada media sosial, sehingga BERT paling di rekomendasikan dalam analisis sentimen MBG.
Kata kunci: analisis sentimen; machine learning; SVM, RF, dan BERT
Abstract:Abstract: The Academic Information System (SIAKAD) plays an important role in supporting the management of academic administration in higher education institutions, particularly for students. ABC University has implemented…
ed SIAKAD since 2018 to facilitate administrative ativities in line with its motto of a high technology campus. This study aims to measure the sucess of SIAKAD usage from the aspects of acceptance, satisfaction, suitability, and perceived benefits. The integration of the Unified Theory of Acceptance and Use of Technology (UTAUT), DeLone & McLean, and Task Technology Fit (TTF) models was carried out to obain a more comprehensive overview in assessing the success of SIAKAD. UTAUT explains the factors influencing the intention to use, DeLone & McLean emphasizes the relationship between system quality and both user satisfaction and net benefits, while TTF evaluates the fit between technology and user tasks. By combining these three models, the study addresses the limitations of each model and produces a more holistic approach in measuring acceptance, success, and the appropriateness of system use. The testing was conducted using SPSS and Structural Equation Modeling (SEM) analysis through AMOS.
Keywords: siakad; utaut; delone&mclean; penerimaan teknologi; sem
Abstract:Abstract: Recommendation systems are becoming increasingly important with the growth of streaming platforms. The purpose of this study is to compare the performance of Content-Based Filtering, Neural Collaborative Filtering,…
ing, and a combination of both in a movie recommendation system. The method used in this study involves retrieving movie details from the TMDB API and ratings from the MovieLens 32M Dataset (2010-2023). Each model's performance is evaluated using evaluation metrics such as RMSE and MAE. The results of this study indicate that Neural Collaborative Filtering achieves the best prediction performance (RMSE = 0.785423, MAE = 0.581262), followed by the hybrid model (RMSE = 0.800863, MAE = 0.660872), while Content-Based Filtering produces low performance and limits the capabilities of the hybrid model. In conclusion, these findings highlight the superiority of latent feature-based models such as NCF that learn directly from user interaction patterns over content-based approaches in the context of modern recommendation systems.
Keywords: content-based filtering; hybrid filtering; movie recommendation; neural collaborative filtering.
Abstrak: Sistem rekomendasi menjadi semakin penting seiring berkembangnya platform streaming. Tujuan dari penelitian ini adalah membandingkan kinerja Content-Based Filtering, Neural Collaborative Filtering dan kombinasi keduanya dalam sistem rekomendasi film. Metode yang digunakan dalam penelitian ini melibatkan pengambilan detail film dari TMDB API dan rating dari dataset MovieLens 32M Dataset (2010-2023). Setiap peforma model dievaluasi dengan menggunakan metrik evaluasi seperti RMSE dan MAE. Hasil dari penelitian ini menunjukkan bahwa Neural Collaborative Filtering mencapai kinerja prediksi terbaik (RMSE = 0.785423, MAE = 0.581262), diikuti oleh model hybrid (RMSE = 0.800863, MAE = 0.660872), sementara Content-Based Filtering menghasilkankan peforma yang rendah dan membatasi kemampuan model hybrid. Kesimpulannya, penelitian ini menyoroti superiotas model berbasis latent feature seperti NCF yang belajar langsung dari pola interaksi pengguna dibandingkan pendekatan berbasis konten dalam konteks sistem rekomendasi modern.
Kata kunci: content-based filtering; hybrid filtering; neural collaborative filtering; rekomendasi film.
Abstract:Abstract: The global transition towards sustainable food systems faces significant challenges in vegetarian food supply chains, including transparency issues, distribution inefficiencies, and quality verification problems.…
s. This research proposes VegeChain development, a decentralized marketplace ecosystem based on smart contracts designed to transform vegetarian food supply chains and accelerate Meatless, Balanced, Green (MBG) program adoption. Using mixed-method methodology integrating blockchain system design, stakeholder analysis, and economic simulation, this research develops a comprehensive technology framework combining blockchain transparency, smart contract automation, and sustainable tokenomics with novel mathematical models. The system implements dynamic pricing algorithms based on Automated Market Maker (AMM) mechanisms, multi-objective optimization for supply chain efficiency, and reputation-based consensus protocols. Simulation results demonstrate that VegeChain implementation can improve supply chain efficiency by 35%, reduce food waste by 28%, and increase consumer trust by 42% measured through validated stakeholder satisfaction surveys (n=456) using 5-point Likert scales with statistical significance p<0.001. Technical innovations include Byzantine Fault Tolerant consensus with 99.9% reliability, gas optimization achieving 67% cost reduction, and real-time quality verification algorithms with 98.7% accuracy.
Keywords: smart contracts; supply chain optimization; automated market makers; blockchain technology; sustainable tokenomics
Abstract:Abstract: This study aims to develop a Deep Learning-based Traffic Flow Detector to automatically and accurately observe traffic flow. Conventional traffic observation is often conducted manually or via CCTV, but it is prone…
rone to human error and difficult to use for real-time trend analysis. In this study, the YOLOv4 method is used to detect four types of vehicles (cars, motorcycles, buses, trucks). To continuously track vehicle movement and address occlusion issues, the Deep SORT algorithm is implemented. The YOLOv4 model used is a pre-trained model and was tested on seven CCTV video recordings obtained from the official website of the Pekanbaru City Transportation Department. The system was implemented on a limited device, the Nvidia Jetson Nano, as a simulation of direct CCTV integration. Test results showed a highest precision of 98%, but the maximum accuracy achieved was only 26%. This low accuracy is influenced by several factors, including video resolution, detection model quality, and lighting conditions. Nevertheless, the system demonstrates potential to support future traffic management and engineering decisions but still requires further optimization, including improving video resolution and quality, retraining the model with a more representative local dataset, using lighter and more accurate detection models, and optimizing the tracking algorithm.
Keywords: deep learning; deepsort; NVIDIA Jetson NANO; traffic flow; YOLOv4
Abstrak: Penelitian ini bertujuan mengembangkan Traffic Flow Detector berbasis Deep Learning untuk mengobservasi arus lalu lintas secara otomatis dan akurat. Observasi lalu lintas konvensional sering dilakukan secara manual atau melalui CCTV, namun rentan terhadap human error dan sulit digunakan untuk menganalisis tren secara real-time. Pada penelitian ini digunakan metode YOLOv4 untuk mendeteksi empat jenis kendaraan (mobil, motor, bus, truk). Untuk melacak pergerakan kendaraan secara berkelanjutan dan mengatasi masalah occlusion, digunakan algoritma Deep SORT. Model YOLOv4 yang digunakan merupakan pre-trained model dan diujikan pada tujuh rekaman video CCTV yang diambil dari situs resmi Dinas Perhubungan Kota Pekanbaru. Sistem ini diimplementasikan pada perangkat terbatas Nvidia Jetson Nano sebagai simulasi penerapan langsung pada CCTV. Hasil pengujian menunjukkan presisi tertinggi mencapai 98%, namun akurasi tertingginya hanya sebesar 26%. Rendahnya akurasi dipengaruhi oleh beberapa faktor seperti resolusi video, kualitas model deteksi, serta kondisi pencahayaan. Meski demikian, sistem ini menunjukkan potensi untuk membantu pengambilan keputusan dalam manajemen dan rekayasa lalu lintas di masa depan, namun masih membutuhkan optimasi lebih lanjut, seperti peningkatan kualitas video input, pelatihan ulang model dengan dataset lokal, penggunaan model deteksi yang lebih ringan dan akurat serta pengoptimalan algoritma pelacakan.
Kata kunci: deep learning deepsort; Nvidia Jetson Nano; traffic flow; YOLOv4