Jurnal Teknik Informatika (Jutif)

STUDENT FOCUS DETECTION USING YOU ONLY LOOK ONCE V5 (YOLOV5) ALGORITHM

Rosalina Rosalina — 2024-05-08

Education has a very important role in life, student involvement in the learning process in the classroom is an important factor in the success of learning. However, some students pay less attention to the lesson, indicating a lack of productivity in learning. The use of machine learning and computer vision techniques has undergone significant development in the last decade and is applied in a variety of applications, including monitoring student attention in the classroom. One of the commonly used techniques in machine learning and computer vision to detect objects is by applying image processing. One of the algorithms implemented for object detection that can provide good results is You Only Look Once. This research proposes the application of YOLOV5 in real time student focus detection and analyzes the performance and computational load of the five YOLOV5 architectures (YOLOV5n, YOLOV5s, YOLOV5m, YOLOV5l, and YOLOV5x) in student surveillance during classroom learning. The dataset used is video data that has been converted into image form, and 297 images are produced. Where, this dataset is divided into 2 classes, namely the "Focus" and "Not Focus" classes. The results show that YOLOV5x has the highest computational load with large parameter values and GFLOPs. However, in term model performance YOLOV5m provides more optimal results than other architectures, with precision of 83.3%, recall of 85.1%, and mAP@50 of 89.9%. The results of this study show that the proposed YOLOV5 model can be a good performing method in detecting student focus in real time.

CYBERBULLYING SENTIMENT ANALYSIS OF INSTAGRAM COMMENTS USING NAÏVE BAYES CLASSIFIER AND K-NEAREST NEIGHBOR ALGORITHM METHODS

Fitri Anisa Nirmala — 2024-05-08

The high number of social media users presents major threats and risks, such as cyberbullying Cyberbullying or cyberbullying is one of the negative impacts of the rapid development of technology and social media. Sentiment Analysis is a technique for extracting text data to obtain information about positive, neutral or negative sentiment. One of Indonesian social media that often gets user sentiment through social media is Instagram. By using the Text Mining technique, the classification method will determine whether a sentiment is positive, neutral or negative. This research uses the Naïve Bayes Classifier (NBC) and K-Nearest Neighbor (KNN) methods with tf-idf weighting accompanied by the addition of an emotional icon (emoticon) conversion feature to determine the existing sentiment classes from tweets about Instagram users. The results of calculations using the first three methods using the Partitionong model, the results using the Naive Bayes method, get an accuracy value of 91.25%, a recall value of 92% and a precision value of 90% and calculations using the KNN method have an accuracy value of 67%, a recall value of 49% and a precision value of 34 %. So it can be concluded that the Naïve Bayes Classifier algorithm has the best performance.

COMPARISON OF NAIVE BAYES AND RANDOM FOREST METHODS IN SENTIMENT ANALYSIS ON THE GETCONTACT APPLICATION

Juan Pala Arisula — 2024-10-20

The rapid growth in the use of social media and instant messaging platform apps has significantly changed the way people communicate. One of the most popular apps is GetContact, a platform focused on identifying the phone numbers of irresponsible people and reducing the impact of spam calls. In cases like this, sentiment analysis is important to understand user responses to the service. In performing sentiment analysis, there are two classification methods that will be used, namely the Naive Bayes and Random Forest methods. This research utilizes the SMOTE technique to handle data imbalance, and the results show that the application of SMOTE successfully improves classification accuracy. The Random Forest model performed better than Naive Bayes, with 80% accuracy, 84% precision, 77% recall, and 80% F1 score for positive sentiments, while Naive Bayes achieved 77% accuracy, 79% precision, 79% recall, and 79% F1 score. Although Random Forest is superior in precision, recall , and F1 score for positive sentiments, it performs almost on par with Naive Bayes in classifying negative sentiments, with 76% precision , 84% recall, and 80% F1 score for Random Forest, and 76% precision, 76% recall , and 76% F1 score for Naive Bayes. This shows that both models provide similar results in identifying negative sentiment overall.

EMPLOYEE VOLUNTARY ATTRITION PREDICTION AT PT.XYZ: ENSEMBLE MACHINE LEARNING APPROACH WITH SOFT VOTING CLASSIFIER

Cagiva Chaedar Bey Lirna — 2024-10-20

This research addresses the complexity of employee attrition challenges at PT.XYZ. The main objective is to develop a predictive system for potential voluntary employee attrition by focusing on an in-depth analysis of the factors contributing to attrition at PT.XYZ. The research utilizes data containing information on the job history of PT.XYZ employees from 2018 to 2023. The method employed in the research is a soft voting ensemble classifier model, incorporating SVM, decision tree, and logistic regression, supported by relevant literature. Analysis and exploration of historical data of PT.XYZ employees are conducted to identify key factors influencing employees' decisions to leave the company. Careful data preprocessing is carried out to ensure dataset quality before applying it to the soft voting classifier model. The results of the soft voting classifier modeling used in this research achieve excellent accuracy in both training and testing datasets with respective accuracy percentages of 99% and 98%. Based on the final results of applying the soft voting classifier model, it is expected to provide deep insights and solutions to enhance employee retention at PT.XYZ.

THE PERCEPTIONS OF SEMARANG FIVE STAR HOTEL TOURISTS WITH SUPPORT VECTOR MACHINE ON GOOGLE REVIEWS

Muhammad Haikal Aufan — 2024-10-20

Travelers on the road sometimes need a hotel to rest. In choosing a hotel, they refer to the ratings or reviews written by users through reviews on Google. This is because not all star hotels provide facilities in accordance with user assessments. This study discusses the analysis of the opinions of tourists who have stayed in 5-star hotels in Semarang through a review of commentary data on Google. The 5-star hotels used as the research are Padma, Gumaya, Tentrem, Grand Candi, Ciputra, and PO. The dataset of the six hotels was obtained through a scraping process then followed by data pre-processing. The data was retrieved from Google Maps using the Chrome Instant Data Scrapper extension. Data preprocessing begins with case folding, tokenizing, filtering, and ends with stemming. Support Vector Machine (SVM) is implemented for sentimen classification process. The results from this study are the majority of 5-star hotel reviews in Semarang tend to have positive rather than negative sentimens. Our model was able to produce an accuracy of 0.87 to 0.98. The highest accuracy was achieved by Ciputra Hotel at 0.98 with 543 positive reviews.

TWITTER (X) SENTIMENT ANALYSIS OF KAMPUS MERDEKA PROGRAM USING SUPPORT VECTOR MACHINE ALGORITHM AND SELECTION FEATURE CHI-SQUARE

Mutiara Sari — 2024-10-20

Ministry of Education, Culture, Research and Technology (Kemendikbudristek) has implemented numerous policies aimed at enhancing the quality of education in the country. One of these policies is Kampus Merdeka program. The program includes various initiatives such as Teaching Campus, the Merdeka Student Exchange program, and Internship and Independent Study programs, which have gained significant popularity among students across Indonesia. However, the Kampus Merdeka program has drawn many pros and cons, with some parties supporting the initiative, but also many criticisms related to its implementation, which is considered not optimal in some educational institutions. Social media is where many of these opinions are voiced, one of the most widely used of which is twitter. In light of these circumstances, this study conducted a sentiment analysis of the independent campus program to assess public sentiment towards it. The dataset used in this research consisted of 500 tweets containing the keyword "kampus merdeka" with 250 tweets reflecting positive sentiment and 250 tweets reflecting negative sentiment. The results of the tests carried out obtained the highest increase in results in the 10:90 ratio, namely with an accuracy that increased by 14% from the previous 66% to 80%, precision also increased by 22% from the previous 67% to 89%, recall increased by 16% from the previous 58% to 79%, and the f1-score value which was previously 62% turned into 79% because it also increased by 17%.

IMPLEMENTATION OF DEEP LEARNING MODELS IN HATE SPEECH DETECTION ON TWITTER USING AN NATURAL LANGUAGE PROCESSING APPROACH

Muhammad Arifin — 2024-10-20

In the digital era, the misuse of the freedom to communicate on the internet often leads to problems such as the spread of hate speech, which can harm individuals based on race, religion, and other characteristics. This issue requires effective solutions for content moderation, particularly on social media platforms like Twitter. This research develops a deep learning model utilizing Natural Language Processing (NLP) to detect hate speech and aims to improve existing content moderation mechanisms. The methods used include data collection, preprocessing through techniques such as case folding, tokenization, lemmatization, and model creation using TensorFlow Extended (TFX) involving embedding, dense, and global pooling layers. The model is trained to optimize accuracy by minimizing the loss function and closely monitoring evaluation metrics. The results show that this model achieves a prediction accuracy of 84%, an AUC value of 0.796, and a binary accuracy of 76%. The conclusion of this research is that the use of deep learning and NLP in detecting hate speech offers a highly potential approach to enhancing digital content moderation, providing a solution that is not only efficient but also accurate.

CLASSIFICATION MODELS FOR ACADEMIC PERFORMANCE: A COMPARATIVE STUDY OF NAÏVE BAYES AND RANDOM FOREST ALGORITHMS IN ANALYZING UNIVERSITY OF LAMPUNG STUDENT GRADES

Dian Kurniasari — 2024-10-20

At the university, students are provided with a comprehensive assessment of their academic achievements for each course completed at the end of every semester. This study aimed to compare the effectiveness of two classification methods, the Naïve Bayes and the Random Forest methods, in classifying student learning outcomes. The research process is segmented into various stages: data selection, data preparation, model building and testing, and model evaluation. The findings indicated that the Naïve Bayes and Random Forest approaches exhibited superior accuracy levels when employing data splitting strategies, in contrast to k-fold cross-validation. Based on the examination, the Random Forest approach demonstrated superiority in identifying the scores of University of Lampung students, achieving an accuracy percentage of 99.38%. Notably, both techniques showed a substantial performance improvement using Gradient Boosting. The Naïve Bayes method attained an accuracy rate of 99.89%, while the Random Forest method reached 99.45%. The results demonstrate that employing the Random Forest classification method consistently leads to superior performance in identifying and classifying student grades. Furthermore, using Gradient Boosting in the boosting process has demonstrated its efficacy in enhancing the classification methods' accuracy. These findings significantly contribute to the comprehension and advancement of evaluation systems for assessing student learning outcomes in the university environment.

THE PERFORMANCE ANALYSIS OF REACTIVE AND PROACTIVE ROUTING PROTOCOLS FOR V2V COMMUNICATION IN DYNAMIC TRAFFIC SIMULATION

Ketut Bayu Yogha Bintoro — 2024-10-20

The research problem addressed in this study arises from the urgent need to enhance Vehicle-to-Vehicle (V2V) communication in dynamic traffic scenarios. V2V communication is a critical component of intelligent transportation systems aimed at improving traffic safety and efficiency. However, existing routing protocols exhibit varying performance under different traffic conditions, such as free flow, steady flow, and congestion. Consequently, a comprehensive comparison is necessary to evaluate the effectiveness of three routing protocols—AODV, LA-AODV, and DSDV—in dynamic V2V scenarios. This research aims to address this problem by simulating realistic traffic conditions and evaluating the Quality of Service (QoS) of each protocol using metrics such as Packet Delivery Ratio (PDR), Packet Loss Ratio (PLR), Throughput, End-to-End Delay, and Jitter. The findings indicate that LA-AODV demonstrates superior performance in terms of PDR (up to 4% at 500 seconds), PLR (reaching 95.33% at 500 seconds), and Throughput (reaching 84.81 Kbps at 800 seconds). This makes it an excellent choice for applications prioritizing reliable data transfer. Conversely, AODV exhibits the lowest latency and jitter, with latency (reaching 7.40E+10 ns) and jitter (reaching 1E+10 ns) at 300 and 400 seconds, respectively. AODV is well-suited for real-time V2V communication due to its minimal delay and jitter. DSDV, while minimizing control overhead, performs less favorably in other metrics. Consequently, AODV emerges as the preferred option for real-time V2V communication. LA-AODV excels in scenarios emphasizing data delivery and high throughput. DSDV may find relevance in security-sensitive applications where minimizing control traffic is crucial.

PAPAYA TYPE CLASSIFICATION USING YOLOv8

Egi Verdiansyah — 2024-10-20

Papaya (Carica papaya L) is a fruit that is easily found in subtropical and tropical regions, including Indonesia. With many varieties of papaya, the manual method used in distinguishing papaya types by humans depends on individual knowledge which can cause inaccuracies in the classification process. The manual classification process also takes a very long time if production is done on a large scale. Therefore, a technology for sorting automation is needed, especially in the industrial world. This research aims to classify papaya classes based on their type. The classification is divided into four classes, namely bangkok papaya, california papaya, hawai papaya, and red lady papaya. The classification process in this study uses the YOLOv8 model, where the total dataset is 1200 papaya images with a training data division of 88% (1050 images), 8% validation data (100 images), and 4% test data (50 images). The dataset is separated according to papaya fruit class. Data training was conducted with 300 epochs. The results show that bangkok papaya has a mAP value of 96%, california papaya 97%, hawai papaya 95%, and red lady papaya has 95% mAP. The average class has a precision value of 99.6%, and recall 100.0%. It can be concluded that the YOLOv8 classification model is able to achieve a high level of accuracy.

CATARACT CLASSIFICATION USING CONVOLUTIONAL NEURAL NETWORK (CNN) INCEPTION RESNETV2

M. Mauludin Zulfa — 2024-10-20

The eye is a human sensory device that functions as an organ of vision. Referring to data from the World Health Organization (WHO) in 2018, cataracts are responsible for 48% of blindness cases in the world and are the main cause in Indonesia. People still find it difficult to distinguish cataract eyes from normal eyes, so they often do not realize the indications of cataract disease. It is important to conduct early detection of cataract disease before blindness occurs. As technology develops, cataract identification becomes easier and simpler with digital image processing classification. This research develops a cataract image classification model using Convolutional Neural Network (CNN) with Inception-ResnetV2 architecture to identify cataract eyes with normal eyes. The proposed model consists of two parts of Inception-ResnetV2 architecture as the base model, and the head model in the form of Fully Connected Layers consisting of global average polling, 2 dense relu layers of 128 and 256 neurons, 2 batch normalization layers, 2 layers of dropout parameter 0.5, and softmax activation function for the output layer. To improve model training, the Stochastic Gradient Descent (SGD) optimization function is used. The dataset consists of 2,192 eye fundus images with 2 main classes of cataract and normal taken from the public data provider site Kaggle. Learning rate tests on the optimization function were carried out with parameters 0.1, 0.01, and 0.001, the results of the proposed model compiled with Stochastic Gradient Descent (SGD) learning rate 0.01 gave a final accuracy of 96%.

CLASSIFICATION OF RICE ELIGIBILITY BASED ON INTACT AND NON-INTACT RICE SHAPES USING YOLO V8-BASED CNN ALGORITHM

Nazwa Putri Hastari — 2024-10-20

The large amount of unfit rice has an impact on the quality of rice provided to the community. This is due to the lack of supervision of the quality of existing rice, so that the quality of rice distributed to the community has a lot of unfit quality. Rice production for public consumption reached 21.69 million tons in 2021, according to data from the Central Statistics Agency (BPS). Rice is the main food of the Indonesian people because most Indonesians are farmers and the vast amount of agricultural land makes Indonesia one of the largest rice producing countries in Southeast Asia, this has a huge impact on people's habits in consuming rice as the main food provider. The Government of the Republic of Indonesia started a Social Assistance rice distribution program through the Ministry of Social Affairs in 2018. This program is named Prosperous Rice Social Assistance (Bansos Rastra). Classification of rice eligibility can be the first step to ensure that the rice received from the government is of high quality and can meet the daily needs of households in Indonesia. CNN algorithm based on YOLOv8 system can automatically recognize the form of rice given by the government whether it is feasible or not. In the research stages there are dataset collection, preprocessing, training models to evaluation. Based on the results obtained in this study, the accuracy achieved is 79% for the Eligible class and 79% for the Ineligible class with Confidence score reaching a value of 1.00. The results of this study can be used as a decent and unfit rice classification detection model by looking at the shape of the rice. So that the rice distributed to the community has decent rice quality.

INFORMATION SYSTEM AUDITING USING COBIT 5 ON PRADITA UNIVERSITY E-LEARNING ASWAYA

Aldira Panduwitama — 2024-10-25

Learning management sytems (LMS) play a crucial role to an academic process in a modern university, LMS systems facilitate an online learning process and help teach and it students to connect. This system is considered vital by its nature to ensure a smooth academic process, that’s why keeping the sensitive data and information that are contained in it are a must and ensuring the security are top priority.Whereheas LMS that are used in Pradita University havent got its system information audited. By using COBIT 5 developed by ISACA gives a solution to align IT with the organization goals, ensure its security, manage risk and threats and also manage compliance to a current policy. The main goal of this research is to understand its IT management, Especially on Security sector and Service request incident that contained in COBIT 5 domains like DSS02, DSS04, DSS05 and how this information system compliance to external regulations that already stated by an external organization as contained in MEA03 domain. And the result of its maturity level that have been assesed on Domain DSS02 is 2 DSS04 level 3 DSS05 level 5 and MEA03 is on level 3. That makes LMS aswaya have an average of 2.5 on its maturity level and have 0,5 gap from the expected level of 3. This shown because most of the processes on LMS Aswaya are Repeatable but inituitive.

ENSEMBLE MACHINE LEARNING WITH NEURAL NETWORK STUNTING PREDICTION AT PURBARATU TASIKMALAYA

Muhammad Al-Husaini — 2024-10-25

This research uses an ensemble model and neural network method that combines several machine learning algorithms used in the prediction of stunting and nutritional status children in Purbaratu Tasikmalaya. This ensemble method is complemented by a combination of the prediction results of several algorithms used to improve accuracy. The data used is anthropometry-based calculations of 195 toddlers with 39% of related stunting from 501 total data in Purbaratu Tasikmalaya City; high rates of stunting this research urgent to make a stable model for prediction. The results of this study are significant as they provide a more accurate and efficient method for predicting stunting and nutritional status in children, which can be crucial for early intervention and prevention strategies in public health and nutrition. The best accuracy value for some of these categories is 98, 21% for the Weight/Age category with the xGBoost algorithm, 97.7% of the best accuracy results with the Random Forest and Decision Tree algorithms for the Height/Age category, the Weight/Height category with the best accuracy of 97.4% for the Random Forest and xGBoost algorithms, and the use of neural network models resulted in an accuracy of 99.19% for Weight/Age and Height/Age while for Weight/Height resulted in an accuracy of 91.94%..

SYSTEMATIC LITERATURE REVIEW OF DOCUMENTS SIMILARITY DETECTION IN THE LEGAL FIELD: TREND, IMPLEMENTATION, OPPORTUNITIES AND CHALLENGE USING THE KITCHENHAM METHOD

Muhammad Furqan Nazuli — 2024-10-25

This research conducted a Systematic Literature Review (SLR) to observe the application of graph mining techniques in detecting document law similarities. Graph mining, where nodes and edges represent entities and relations respectively, has proven effective in identifying patterns within legal documents. This review encompasses 93 relevant studies published over the past five years. Despite its potential, graph mining in the legal domain faces challenges, such as the complexity of implementation and the necessity for high-quality data. There is a need to better understand how these techniques can be optimized and applied effectively to address these challenges. This SLR utilized a comprehensive approach to identify and analyze trends, implementations, and popular domains related to graph mining in legal documents. The study reviewed trends in the number of studies, categorized the implementations, and evaluated the prevalent techniques employed. The review reveals a growing trend in the use of graph mining techniques, with a noticeable increase in the number of studies year by year. The implementation of these techniques is the most popular category, with applications predominantly in legal domains such as laws, legal documents, and case law. The most frequently used graph mining techniques involve Natural Language Processing (NLP), Information Retrieval, and Deep Learning. Although challenges persist, including complex implementation and the need for quality data, graph mining remains a promising approach for developing future information systems in law.

IMPLEMENTATION OF REST API ARCHITECTURE FOR FEELSQUEST ONLINE COURSE FEATURE IN FEELSBOX APPLICATION USING LARAVEL FRAMEWORK

Faza Alexander Riawan — 2024-10-25

Feelsbox is a digital-based startup that focuses on the importance of mental health issues and offers innovative solutions to help people maintain their mental health. FeelsBox took the initiative to develop an online course feature "FeelsQuest" with the aim of providing education and helping prevent and overcome mental health problems to the wider community, especially teenagers. The development of this feature uses the PHP programming language with the Laravel framework and implements the REST API architecture. The choice of REST API architecture is based on the concept of separation of responsibilities so that the API can be reused on different platforms. In addition, a suitable test is needed to test the REST API that has been built. Testing of the REST API that has been built is done with the API testing method which is focused on aspects of functionality and performance using Postman to ensure that the API built produces responses and behaves according to the needs of the FeelsQuest feature of the FeelsBox application. The test results show that the implementation of the REST API on the FeelsQuest feature is in accordance with the functional requirements and successfully applies the concept of separation of concerns and meets the non-functional needs of the FeelsQuest feature related to the response time of each API, which is under 3 seconds.

DEVELOPMENT OF MOBILE-BASED FREELANCE SERVICES MARKETPLACE WITH FEATURE-DRIVEN DEVELOPMENT METHODOLOGY

Mochamad Ikhsan Nurdiansyah — 2024-10-25

The rapid advancement of technology demands that humans adapt to the evolving workplace. One of the impacts of technological development is the emergence of freelancers who use online platforms. Online platforms make it easier for freelancers and clients to collaborate. Currently, there are various Freelance Services Marketplace platforms available. However, it is important to note that the developers of these platforms are from outside Indonesia, which poses a risk to the security of people's data and the ease of payment applicable in Indonesia. With the high number of mobile device users in Indonesia,, the development of a mobile-based Freelance Services Marketplace application system that is tailored to local needs and takes into account the data security of Indonesian people, as well as involving local developers, is needed. This research aims to overcome these problems by designing a mobile-based freelance services marketplace application system using the Feature-Driven Development (FDD) method. The selection of FDD as an application development methodology is based on complex feature requirements, good planning management, emphasis on feature quality, and structured. The development process follows the FDD approach starting from feature planning and design, scheduling, implementation, and testing. The application design in this research is based on user needs with a focus on the features needed. This solution is important because it can be used by clients to find freelancers who match their needs and provide opportunities for freelancers to offer their services properly with guaranteed data security and transactions. The main result of this research is that the FDD method can help in the development of a freelance service marketplace application by paying attention to the systematic or structured level, quality, and security of the application.

IMPLEMENTATION OF A COMBINATION OF ADVANCED ENCRYPTION STANDARD CRYPTOGRAPHY WITH SUBBYTES MODIFICATION AND STEGANOGRAPHY BASED ON A WEBSITE

Muhammad Ilham Kurniawan — 2024-10-25

The Advanced Encryption Standard (AES) is a symmetric encryption algorithm commonly used to protect digital data. However, concerns about potential attacks on cryptographic keys and the development of cryptanalysis methods further reinforce the need for security enhancement. This study aims to combine two technologies: the Advanced Encryption Standard (AES) cryptography with modifications to the SubBytes, and steganography using the Least Significant Bit (LSB) method in images, to enhance the security level of encrypted messages in the context of transmission through websites. In this study, modifications were made to the AES algorithm by replacing the S-box in the SubBytes process with a perfect SAC S-box with an average SAC value of 0.5. This testing is divided into two types: algorithm testing and system testing. Algorithm testing involves performance testing methods that show longer decryption times with an average difference of 80.27 milliseconds, cryptanalysis testing showing increased ciphertext security based on cryptanalysis time estimates using brute force, and randomness testing to demonstrate improvements in Frequency and Poker tests. System testing using the Black Box method shows results that are valid as expected.

SENTIMENT ANALYSIS FOR E-COMMERCE PRODUCT REVIEWS BASED ON FEATURE FUSION AND BIDIRECTIONAL LONG SHORT-TERM MEMORY

Habibullah Akbar — 2024-10-25

E-commerce platforms would benefit from performing sentiment analysis of their customer's feedback. However, the vast amount of transaction data makes manual sentiment analysis of product reviews impractical. This research proposes an approach to automatically classify the sentiment of a given product review based on three major steps: data preprocessing, text representation, and classification model development. First, review data is cleaned to remove ambiguity and non-meaningful elements. Second, Word2Vec and GloVe features are combined to represent the words in a more unified vector space. Lastly, these combined features are classified to determine sentiment polarity using the Bidirectional Long Short-Term Memory Network (BiLSTM) model. The test results demonstrate that the proposed BiLSTM model achieves 91% uniform performance for all four metrics (accuracy, precision, recall, and F1-score), which is 3% higher than the results achieved by the standard LSTM model. Moreover, the BiLSTM model requires 9.91 seconds less training computation time than the LSTM.

CHATBOT FEATURES ON WEBSITES USING DIALOGFLOW FRAMEWORK WITH RULE-BASED METHOD

Muhammad Afiq Nabiha Riandika — 2024-10-25

A chatbot is an artificial intelligence (AI) technology that can mimic human conversation in the form of text or voice messages through a website, or mobile application. Chatbots are widely used to facilitate communication, such as finding information, or services. In this research, the difficulty of accessing information in obtaining answers to questions asked by the public, as well as taking too long for the admin to reply when providing information to people who ask questions related to information about the topic raised, is an initial problem that will be solved in this research. Chatbot is a solution that can overcome the above problems, chatbot itself is designed to help food license applicant services on the website of the Semarang City Health Office in the field of Pharmacy and Perbekes. By utilizing the Dialogflow framework, this chatbot will use the Rule-Based method because in this development, the Rule-Based method can adjust common questions and answers that are often asked by the public, and can also be changed and even expanded to manage conversations without experiencing much difficulty in changing them that follow questions that are often asked over time. The system consists of agents, intentions, and training phrases that will be trained to understand various questions and provide relevant responses. This chatbot development aims to improve the efficiency of food licensing services, reduce applicant waiting time, and provide accurate and easily accessible information. The test results in this development are on a chatbot system that can run well, and is able to understand various kinds of questions related to food licensing, and provide appropriate responses in accordance with the predetermined intent. In addition, an evaluation of the level of user satisfaction will be carried out to measure the success of this system. This chatbot can improve the quality of public services in the field of food licensing and provide convenience for the public in processing licenses.

HORTICULTURE SMART FARMING FOR ENHANCED EFFICIENCY IN INDUSTRY 4.0 PERFORMANCE

Nurhikma Arifin — 2024-10-25

Chili peppers and papayas are important horticultural commodities in Indonesia with high economic value. To enhance productivity and efficiency in cultivating these crops, the application of Smart Farming technology is crucial. This study evaluates the use of image processing and artificial intelligence in the pre-harvest and post-harvest processes for chili peppers and papayas. For the pre-harvest process, data from 50 images of ripe chili peppers on the plant were used. The counting of ripe chilies was performed using HSV color segmentation with two masking processes, resulting in an average accuracy of 82.58%. In the post-harvest phase, 30 images of papayas, consisting of 10 images for each ripeness category—unripe, half-ripe, and ripe—were used. Papaya ripeness classification was carried out using the Support Vector Machine (SVM) algorithm with a Radial Basis Function (RBF) kernel and parameters C = 10 and γ = 10^-3, achieving perfect classification accuracy of 100% for all categories. This study underscores the significant potential of Industry 4.0 technologies in enhancing agricultural practices and efficiency in the horticultural sector, providing important contributions to optimizing chili pepper and papaya production.

CLASSIFICATION OF FAMILY HOPE PROGRAM RECIPIENTS USING NAIVE BAYES AND C4.5 METHODS

Farras Ahmad Fauzi — 2024-10-25

Receiving PKH assistance in Rawamerta District does not always go well, so there are people who are not entitled to receive assistance. This is because there is still no system that can facilitate the process of classifying PKH assistance recipients. The application of data mining can facilitate classification with high speed and accuracy. The purpose of this study is to classify PKH assistance recipients using the Naïve Bayes and C4.5 methods to determine the eligibility of PKH for people facing social welfare problems. The data used is PKH data in Rawamerta District, Karawang Regency in 2023, totaling 1834 data. The results of naive bayes accuracy of 98.89%, precision 98.25%, recall 98.51%, F1-score 98.89%, and AUC 1.00 are included in the excellent classification because they are in the range of 0.90-1.00, while the C4.5 algorithm produces Accuracy values of 99.26%, Precision 99.25%, Recall 99.25%, F1-score 99.25% and AUC 0.99 are included in the excellent classification because they are in the range of 0.90-1.00. The C4.5 algorithm is superior to Naive Bayes, because the accuracy produced is higher.

FACIAL PHOTO AUTHENTICITY DETECTION USING FACE RECOGNITION AND LIVENESS DETECTION

Bimo Vallentino Achmad — 2024-10-27

Facial recognition has been widely adopted by many systems as authentication. However, relying on facial photos for authentication is insufficient, as these can be manipulated using printed or digital photos. One method that can be used to prevent this is to integrate face recognition with liveness detection. In this research, face recognition and liveness detection are implemented using a Convolutional Neural Network (CNN) because CNN has the ability to process and extract features from photos effectively. There are two types of datasets used, namely CelebA-Spoof for liveness detection and lfw-deepfunneled for face recognition. The face recognition model achieved good accuracy and loss results of 0.9153 and 0.0514, very promising. Meanwhile, the liveness detection accuracy and loss were 0.8633 and 0.7166.

COMPARISON OF K-NEAREST NEIGHBORS AND NAÏVE BAYES CLASSIFIER ALGORITHMS IN SENTIMENT ANALYSIS OF USER REVIEWS FOR INTERMITTENT FASTING APPLICATIONS

Muhammad Varhan Kusuma — 2024-10-28

Applications that focus on health, especially obesity prevention, are scattered in the Google Play Store, one of which is the "Intermittent Fasting" application, which, according to the developer, aims to help users maintain a healthy lifestyle and regulate eating habits. With the increasing number of similar health applications, this research focuses on sentiment analysis of user reviews of "Intermittent Fasting" to find out how users respond. The purpose of this research is to find the best algorithm to analyze sentiment on user reviews on the Google Play Store against the "Intermittent Fasting" application, as well as provide recommendations for new or old users and for application developers based on the results of processing review data. The data mining methodology used in this research is CRISP-DM, using a dataset collected on user reviews on the Google Play Store for five years (2019-2024), which is annotated with three sentiment labels (positive, negative, and neutral) based on user ratings, then modeling using two algorithms K-Nearest Neighbors (KNN) and Naïve Bayes Classifier (NBC). The contribution of this research is to test, evaluate, and compare the two algorithms (KNN and NBC) using two testing models (Split and K-Fold Cross Validation) and then provide recommendations for the best algorithm. The research concludes that the NBC algorithm is superior to KNN with an accuracy value of 80%, while the KNN algorithm has an accuracy value of only 71.43%. In addition, the K-Fold Cross Validation testing model is more optimal in improving the accuracy of the algorithm's performance than the Split model.

LEARNING RATE AND EPOCH OPTIMIZATION IN THE FINE-TUNING PROCESS FOR INDOBERT’S PERFORMANCE ON SENTIMENT ANALYSIS OF MYTELKOMSEL APP REVIEWS

Muhammad Naufal Zaidan — 2024-10-28

With the advancement of the digital era, the growth of mobile applications in Indonesia is rapidly increasing, particularly with the MyTelkomsel app, one of the leading applications with over 100 million downloads. Given the large number of downloads, user reviews become crucial for improving the quality of services and products. This study proposes a sentiment analysis approach utilizing the Indonesian language model, IndoBERT. The main focus is on optimizing the learning rate and epochs during the fine-tuning process to enhance the performance of sentiment analysis on MyTelkomsel app reviews. The IndoBERT model, trained with the Indo4B dataset, is the ideal choice due to its proven capabilities in Indonesian text classification tasks. The BERT architecture provides contextual and extensive word vector representations, opening opportunities for more accurate sentiment analysis. This study emphasizes the implementation of fine-tuning with the goal of improving the model's accuracy and efficiency. The test results show that the model achieves a high accuracy of 96% with hyperparameters of batch size 16, learning rate 1e-6, and 3 epochs. The optimization of the learning rate and epoch values is key to refining the model. These results provide in-depth insights into user sentiment towards the MyTelkomsel app and practical guidance on using the IndoBERT model for sentiment analysis on Indonesian language reviews.

IMPLEMENTATION OF DEEP LEARNING FOR DETECTING PHISHING ATTACKS ON WEBSITES WITH COMBINATION OF CNN AND LSTM

Ahmad Raihan — 2024-10-28

Phishing attacks represent significant cyber threats to internet users, particularly on websites. These attacks are conducted by perpetrators seeking to acquire victims' data by impersonating legitimate websites. To address this threat, a solution is proposed using deep learning with a combined algorithm of convolutional neural network and long short-term memory. The research methodology included data collection comprising phishing and legitimate website links, pre-processing through tokenization, padding, and labeling, and splitting data into training and testing sets. The models were then trained, and grid search was employed to identify the optimal hyperparameters for each algorithm. The algorithm’s performance was calculated by accuracy, precision, recall, and F1-score metrics. The outcomes indicated that using the combination algorithm achieved 95.63% accuracy, 94.60% precision, 96.81% recall, and 95.78% f1-score. This paper concludes the proposed algorithm is effective in detecting phishing attacks on websites.

WEB-BASED IMAGE CAPTIONING FOR IMAGES OF TOURIST ATTRACTIONS IN PURBALINGGA USING TRANSFORMER ARCHITECTURE AND TEXT-TO-SPEECH

Safa Muazam — 2024-10-29

Purbalingga is a region located in Central Java Province, offering interesting natural beauty and tourist destinations. Many tourists capture their moments in photos, which are then uploaded to social media. However, a picture can contain a lot of information, and each individual may interpret it differently. Without captions, people may struggle to extract this information. Image captioning addresses this challenge by automatically generating text descriptions for images. Additionally, text-to-speech is used to enhance accessibility for the visually impaired in understanding image descriptions. This research aims to develop an image captioning model for images of tourist attractions in Purbalingga using transformer architecture and ResNet50. The transformer architecture employs an attention mechanism to learn the context and relationships between inputs and outputs, while ResNet50 is a robust convolutional network for image feature extraction. Model evaluation using BLEU metrics, which compare generated sentences to reference sentences, shows the best results as BLEU-{1, 2, 3, 4} = {0.672, 0.559, 0.489, 0.437}. Experiments indicate that increasing embeddings and layers extends training time and lowers BLEU scores, while changing the number of heads has minimal impact on results. The best model is implemented in a web-based application using the SDLC waterfall method, Flask framework, and MySQL database. This application allows users to upload tourist attraction images, receive automatic descriptions in Indonesian, and listen to the captions read aloud using the Web Speech API-based text-to-speech feature. Blackbox testing results show valid outcomes for all tests, indicating that the application operates as required and is suitable for use.

SPEARMAN CORRELATION ANALYSIS OF AIR AND BILLET TEMPERATURE IN ALUMINUM HOMOGENIZATION USING IoT-BASED REAL-TIME DATA COLLECTION

Filzah Amanina Afiqah — 2024-10-30

Understanding the correlation between the parameters involved in the homogenization process of aluminum billets contributes to better process control. As a critical step in the production of aluminum billets, failure to control the homogenization temperature can lead to variations in product quality and a negative effect on mechanical strength. To address this issue, this study aims to understand the correlation between air and billet temperature of homogenization using temperature data obtained from thermocouple sensors placed at different points in the oven and billet. The temperature data was collected in real time through an Internet of Things (IoT) network. Spearman correlation analysis was performed on the collected data to determine the relationship between temperatures at different measurement points. The analysis results show that the air temperature at the Z2 right point had a strong correlation with the billet temperature, with a correlation value of 0.90. In contrast, the correlation between air temperature and billet temperature at Z1 Left was lower, indicating a weaker correlation and resulting in uneven heat distribution. These results highlight the importance of controlling the air temperature at Z2 Right to improve the temperature distribution during heat treatment. In addition, this study provides a real case in the implementation of real-time monitoring technology for better understanding on industrial process, especially heat treatment process.

IDENTIFYING POTENTIAL CREDIT CARD PAYMENT DEFAULTS USING GMDKNN WITH LOF AS OUTLIER HANDLING

Liony Puspita Dewi — 2024-11-02

In classifying data, accuracy results are greatly influenced by outliers. The presence of outliers can cause a low level of accuracy in the classification process. The Generalised Mean Distance K-Nearest Neighbor (GMD-KNN) algorithm is a classification technique that shows advantages in terms of flexibility and responsiveness to attribute variations. This research aims to classify credit card data between current and bad payments by handling outliers using the Local Outlier Factor (LOF). The data used is 30,000 credit card transaction data taken from the UCI Machine Learning Repository. This research method uses several stages, namely data collection, data pre-processing carried out to detect and clean outliers with LOF, classification process with GMD-KNN, and evaluation to calculate the accuracy of classification results. As a result, the model shows the best performance at 80%:20% data sharing ratio with k=5 value, achieving 77.60% accuracy, 74.97% precision, 82.57% recall, 78.58% F1-Score, and 77.48% G-Mean.