8th Joint Symposium on Computational Intelligence (JSCI8)

Assoc. Prof. Dr. Phayung Meesad

Department of Information Technology Management; Dean, Faculty of Information Technology and Digital Innovation, KMUTNB, Thailand

Dr. Chalermpol Tapsai, College of Innovation and Management, Suan Sunandha Rajabhat University, Thailand

Abstract: Artificial Intelligence is now a big challenge. In Thai Natural Language Processing (NLP) is one of the most problems in AI Thai community. This talk will present how to solve Thai NLP problems. Firstly, Natural Language Processing for Data Retrieval and Information Processing (NLP-DRIP) will be presented. In addition, several implemented algorithms that can improve both lexical analysis and semantic analysis processes. Ranking Trie is an effective algorithm for improving word segmentation that uses words’ frequency to rearrange Trie’s structure and reduce the size of the dictionary. In addition, Completed Soundex, a phonetic coding and similarity parsing algorithm for typos error recovering and Pattern Parsing and Ontology used for semantic analysis in order to allow users to retrieve and process data with various patterns of sentences and conditions are presented.

Assoc. Prof. Dr. Akhilesh Kumar Sharma, Department of Information Technology, Manipal University Jaipur (Rajasthan), India

Abstract: Nature-Inspired Optimization and the related algorithms are the innovative mathematical methods for non-linear optimization that have their origins in the way various species behave in order to optimize their chances of survival. Nature-inspired algorithms emulate processes that are found in the natural world. Nature gives opportunities based on that many complex problems can be solved and optimized so that the solutions can be obtained. This talk provides a quick reference to a few popular nature-inspired algorithms for optimization.

Saranlita Chotirat and Phayung Meesad

Abstract:  The purpose of this study was to study question classification that automates define the category of wh-questions from Thai text. This was achieved compared efficiently classify text to a wh – question class of simple sentences and interrogative sentence trained through natural language processing (NLP) which considerate the top 5 and top 10 POS (Part of Speech Tagging) and used classification models (Naïve Bayes, Logistic Regression, Support Vector Machine, K-Nearest Neighbors, and neural network which employs long short-term memory (LSTM) to classify the question categories. The experimental results showed that the average classification accuracy up to 78.00%, precision 81.76%, recall 78.00% and F1 score 79.84%,which suggests that filter words only Top5 POS can solve the problem of classifying Thai text to effectively question classification.

Wanthanee Rathasamuth and Kitsuchart Pasupa

Abstract: In principle, a porcine Single Nucleotide Polymorphism (SNP—a specific piece of nucleotide in a DNA sequence) can be associated with a trait of an individual pig, like its meat quality or resistance to common diseases. It is most desirable to obtain a smallest number of most significant SNPs in genomic research and several computer classification algorithms have been used to find a small number of SNPs. This study proposed a vertically distributed feature selection method incorporating a modified binary flower pollination and a support vector machine classifier for selecting significant porcine SNPs. The proposed method was evaluated and compared against four baseline methods. It provided a mean number of 128.4 selected SNPs that resulted in 94.57% classification accuracy.

Sukanchalika Boonmatham and Phayung Meesad

Abstract: Finding a relationship between corporate news and the stock price is a challenging task. In this research, we try to build machine learning models that capture the relationship of news and stock prices of several companies. In this work, eight companies were selected randomly from Industry Group Index and Sectoral Index. Corporate news articles from the eight selected companies were collected along with their stock prices. Two of traditional machine learning models and two deep learning models were used in this study for comparison purpose. The models were based on Support Vector Machine (SVM), Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). Using news articles as inputs, the models were trained to classify stock prices into two classes: Up and Down of the stock closing price. For classification performance, Accuracy, Precision, and Recall were used. The results showed that GRU had highest precision and recall values with 0.67 and 0.62, respectively. GRU model also had an average accuracy score higher than other models with 0.65.

Professor Dr.-ing.habil. Herwig Unger, Chair of Communication Network, Faculty of Computer Science and Mathematics, Fern University in Hagen, Germany

Abstract: Expert and recommender systems are usually basing on the use of well selected knowledge of specialists, which must be prepared and kept in computing systems in cooperation with computer engineers. This process may require significant time and can be the source of mistakes and incompleteness. The presented new approach intends to use instead of experts the huge amount of available written documents in the desired fields and suggests their automatic processing by our newly developed tools. This allows the processing of a so far hard to reach amount of different sources. An interactive inference method basing on the use of text representing centroids and a fast processing using the method of spreading activation in combination with a new interface guarantee a reliable answering of inquiries formulated by keywords or questions. A presentation of the just finished, new medical recommender will conclude the talk together with a short introduction to TheBrain project.

Carlos Estevão Bastos Sousa, Cláudio Marques De Sá Medeiros, Renato Francisco Pereira, Mateus Alves Vieira Neto and Alcides Andrade Neto

Abstract: Tanneries acquire hides, in most cases, from rural workers, so, due to the informality of the creation, slaughter and extraction of the animal’s skin, they receive them with different types and different defect levels. That said, classifying acquired and processed skins become very complex and tiring activity. The leather discrimination process is completely handmade and subjective, too dependent on the experience of the professional responsible for this step, which, due to tiredness, stress and other factors, end up generating several errors in this process. Currently, there are several studies in the literature related to the detection of leather flaws, however, few studies go further and qualify the skins based on the detected problems. In view of this factor, a system based on Computer Vision and Artificial Intelligence is proposed in which it obtains an accuracy rate of 95.9 % in the detection of faults in wet blue goatskin and 93.0 % in the identification of the quality level of these parts.

Phyu Mar Kyu and Kuntpong Woraratpanya

Abstract: Nowadays, the proliferation of automobile industries is directly related to the number of claims in insurance companies. So insurance companies are in contact with many claims and face solving claims leakage problem, which causes huge losses by ineffective claims processing, frauds, and poor decision making. Advance in Artificial Intelligence (AI), machine learning and deep learning algorithms can help to solve these kinds of problem for insurance industries. In this paper, we aim automated to apply deep learning-based algorithms for damage detection and assessment in the real world. An objective of this paper is automatically to detect the damaged part of a car, assess its location and severity. Initially, we discover the effect of domain-specific pre-trained CNN models, which are trained on ImageNet dataset, followed by fine-tuning because some of the categories can be fine-granular to get our specific task. Then we experiment transfer learning in pre-trained VGG models and using some techniques to improve the accuracy of our system. After analysing our models, we find out the results of using transfer learning and regularization work better than the results of fine-tuning. We get the accuracy of 94.56% and 94.35% in the damaged detection, the accuracy of 74.39% and 76.48% in damage localization, the accuracy of 54.8% and 58.48% in damage severity in VGG16 and VGG19 models by using both of transfer learning and regularization. Actually, we don’t have enough time to train VGG19 model like VGG16. If we can train VGG19 like VGG16, we are sure to get a better result than VGG16.

Risul Islam Rasel, Anupam Chowdhury, Meherab Hossain and Phayung Meesad

Abstract: Heart block occurs when the flow of electricity interrupted or partially delayed between the top chamber and bottom chamber of the heart. People are now more often affecting by this kind of disease. However, early prediction of heart block can reduce the diagnosis complexity and treatment cost. In this study, a data mining and machine learning model is proposed to predict three types of heart blocks, such as 1st degree A-V block, Left Bundle Branch Block (LBBB), and Right Bundle Branch Block (RBBB). Experiment data samples are collected from the cardiology department of Chittagong Medical College Hospital (CMCH), Bangladesh. The dataset contains 32 types of numeric and categorical features about the patient’s ECG report, daily activities, and food habits. The prediction model is designed, trained, and tested with some empirical machine learning algorithms namely Decision Tree, Random Forest, K-Nearest Neighbor, and Support Vector Machine. Finally, the experimentation shows that Decision Tree and Random Forest models outperform the other algorithms in overall heart block prediction with an accuracy of more than 92%.

Panida Lorwongtrakool and Phayung Meesad

Abstract: Contamination inspection or quality inspection of raw materials or products is a very important task, especially in the perfume industry that requires an expert for inspection. However, the human nose has limitations such as fatigue, which affects the accuracy. Therefore, an electronic nose or sensor array has been developed to assist in the inspection. The signal data from electronic noses is fed into machine learning models to learn and process. Since data change over time, the input data fluctuate according to the changing environment. In addition, when there are new data with features that change from the original patterns, the classification outcome may not correct and the model will not be able to classify as effective as the original model. Therefore, to solve the problem mentioned this research presents an incremental learning algorithm based on correlative measurement, called Correlation-Based Incremental Learning Network (CILN) combined with Sliding Window, which learns automatically by adapting to new data while keeping the existing knowledge. The experiments were conducted on classifying perfumes. The experimental data were divided into 50% for training and 50% for testing. The proposed algorithm was compared with other well-known classifiers. The results showed that the proposed CILN algorithm provides the highest accuracy of 100%.