Journal of Research in Artificial Intelligence for Systems and Applications

Banana Pest And Disease Expert System Using Forward Chaining and Certainty Factor

Muhammad Ariful Furqon, Lazarus Dwi Poertantono, Nova El Maidah — Tue, 07 Jan 2025 00:00:00 +0700

Farmers do not understand the various forms of banana plant illnesses. In addition, the inadequate guidance offered by agricultural instructors to banana farmers also leads to issues such as crop failure. Considering these issues requires an expert system capable of aiding farmers in identifying and diagnosing banana pests and diseases. The inference techniques implemented in constructing this expert system are forward chaining and certainty factor. The research process comprises the following phases: (1) acquisition of data; (2) forward chaining inference; (3) weighting of symptoms; and (4) development of expert systems. The banana’s pests and diseases expert system has yielded ten distinct diseases and pests, each characterized by 37 symptoms. Data processing employs forward chaining to generate decision trees and perform pruning to establish criteria for determining symptom statements. The combined certainty computation yields a percentage value of 97.25875968% with a significant confidence level

Comparative Analysis of Machine Learning Algorithms with Sequential Feature Selection and Gridsearch Optimization

Tasya Oktaviana Dwi Cahyanti, Nelly Oktavia Adiwijaya, Gama Wisnu Fajariyanto — Tue, 07 Jan 2025 00:00:00 +0700

Hepatitis is a dangerous disease that can cause liver damage. It is often difficult to diagnose because the symptoms of different categories of hepatitis are hard to distinguish. Hepatitis that is not promptly and properly managed, especially in individuals with chronic liver conditions, can lead to complications. This study aims to improve model performance in hepatitis diagnosis using medical data from hepatitis A and B patients at Citra Husada Hospital in Jember. The results show that the method of gridsearch tuning with SMOTE and SFS is very effective in enhancing the performance of the random forest and extra tree algorithms in managing hepatitis symptom data. The best performance was achieved by the random forest algorithm with an 80:20 data ratio, reaching the highest accuracy of 94.87%, recall of 95.96%, precision of 93.33%, f1-score of 94.07%, specificity of 97.97%, and ROC AUC of 99.13%

Combining Rule Based, Lexicon Based and Support Vector Machine for Improve Accuracy in Sentiment Analysis of ChatGPT Usage

Arham Zainul Abidin , Muhammad Ariful Furqon, Saiful Bukhori — Tue, 07 Jan 2025 00:00:00 +0700

ChatGPT is one of the Large Language Models (LLM) which is an artificial intelligence (AI) based chatbot. ChatGPT caused controversy in various news media due to its ability to interact and provide natural, human-like responses. This controversy created skepticism in society regarding the brand image. Therefore, a sentiment analysis was carried out specifically targeting Indonesian speaking Twitter users with a focus on the ChatGPT brand image. Various studies have been carried out to analyze user sentiment towards ChatGPT by analyzing tweets shared regarding ChatGPT using machine learning and deep learning. This research uses a combination of rule-based, Support Vector Machine (SVM), and lexicon-based methods. Rule-based classification uses emoticons and comparatives, while lexicon-based classification uses the BabelSenticNet lexicon. The data set, obtained through a Twitter crawl, initially consisted of 2,500 tweets, which were then cleaned to 1,728 tweets. After classification, model evaluation is carried out by comparing the performance between the proposed classification model and the constituent models. The proposed classification model has better performance, achieving 86.3% accuracy, 87.1% precision, 86.5% recall, and 86.6% f-measure. Sentiment predictions from 1,728 tweets resulted in 739 positive, 385 negative and 604 neutral sentiments. In conclusion, ChatGPT's brand image in Indonesia tends to be positive, even though there are differences in views and objective assessments.

Detection of Infectious Diseases in Broiler Chickens Based on Chicken Feces Images using A Modified YOLOv5 Algorithm

Anggi, Dwiretno Istiyadi Swasono, Arief — Tue, 07 Jan 2025 00:00:00 +0700

This research focuses on detecting infectious diseases in broiler chickens through feces images using a modified YOLOv5 algorithm. Early disease detection is crucial for maintaining product quality and protecting consumer health. YOLOv5’s high detection speed and accuracy make it suitable for identifying multiple objects, enabling farmers to anticipate and mitigate disease risks. The study used a dataset of 8,067 images across four categories: coccidiosis, healthy, NCD, and salmonella, split into 90:10 and 80:20 training-validation ratios. Six modified YOLOv5 models were tested. Key modifications included adding bottleneck layers, replacing SiLU activation with ELU, and incorporating CBAM in various configurations. The best result, with a 90.0% mAP, came from the sixth modification (C3CBAM layer) with a 90:10 split. Testing via a mobile application showed the second (80:20 split) and fourth (90:10 split) models excelled in detecting both large and small objects.

Optimization of Coffee Bean Maturity Classification by Segmentation on Multispectral Images Using HSV and DBSCAN

Tue, 07 Jan 2025 00:00:00 +0700

In the coffee industry, sorting the maturity level of coffee beans is still done conventionally. In an effort to get good quality coffee beans, automatic classification of the maturity level of coffee beans is needed. The data in this research is multispectral image data and still has a background, so the preprocessing process is the main focus in this research to improve the performance of segmentation analysis in identifying objects and background image data. In the image data of 15 types of channels, a combination of 3 channel variations is carried out by applying HSV transformation so that the image data is easily processed by a computer, then the image data will be clustered using DBSCAN to identify coffee bean objects. The results obtained, the best channel combination in segmentation is blue, azure and amber, namely with a final weight value of 611. The segmentation results in the image data preprocessing process resulted in 100% accuracy. Meanwhile, the performance of the model without the segmentation preprocessing stage resulted in an accuracy of 92%. In conclusion, the performance of the model will be more optimal if preprocessing is done, namely segmentation in separating object and background data.