• Journal of Intelligence and Information Systems
• /
• v.26 no.1
• /
• pp.97-117
• /
• 2020

Product evaluation criteria is an indicator describing attributes or values of products, which enable users or manufacturers measure and understand the products. When companies analyze their products or compare them with competitors, appropriate criteria must be selected for objective evaluation. The criteria should show the features of products that consumers considered when they purchased, used and evaluated the products. However, current evaluation criteria do not reflect different consumers' opinion from product to product. Previous studies tried to used online reviews from e-commerce sites that reflect consumer opinions to extract the features and topics of products and use them as evaluation criteria. However, there is still a limit that they produce irrelevant criteria to products due to extracted or improper words are not refined. To overcome this limitation, this research suggests LDA-k-NN model which extracts possible criteria words from online reviews by using LDA and refines them with k-nearest neighbor. Proposed approach starts with preparation phase, which is constructed with 6 steps. At first, it collects review data from e-commerce websites. Most e-commerce websites classify their selling items by high-level, middle-level, and low-level categories. Review data for preparation phase are gathered from each middle-level category and collapsed later, which is to present single high-level category. Next, nouns, adjectives, adverbs, and verbs are extracted from reviews by getting part of speech information using morpheme analysis module. After preprocessing, words per each topic from review are shown with LDA and only nouns in topic words are chosen as potential words for criteria. Then, words are tagged based on possibility of criteria for each middle-level category. Next, every tagged word is vectorized by pre-trained word embedding model. Finally, k-nearest neighbor case-based approach is used to classify each word with tags. After setting up preparation phase, criteria extraction phase is conducted with low-level categories. This phase starts with crawling reviews in the corresponding low-level category. Same preprocessing as preparation phase is conducted using morpheme analysis module and LDA. Possible criteria words are extracted by getting nouns from the data and vectorized by pre-trained word embedding model. Finally, evaluation criteria are extracted by refining possible criteria words using k-nearest neighbor approach and reference proportion of each word in the words set. To evaluate the performance of the proposed model, an experiment was conducted with review on '11st', one of the biggest e-commerce companies in Korea. Review data were from 'Electronics/Digital' section, one of high-level categories in 11st. For performance evaluation of suggested model, three other models were used for comparing with the suggested model; actual criteria of 11st, a model that extracts nouns by morpheme analysis module and refines them according to word frequency, and a model that extracts nouns from LDA topics and refines them by word frequency. The performance evaluation was set to predict evaluation criteria of 10 low-level categories with the suggested model and 3 models above. Criteria words extracted from each model were combined into a single words set and it was used for survey questionnaires. In the survey, respondents chose every item they consider as appropriate criteria for each category. Each model got its score when chosen words were extracted from that model. The suggested model had higher scores than other models in 8 out of 10 low-level categories. By conducting paired t-tests on scores of each model, we confirmed that the suggested model shows better performance in 26 tests out of 30. In addition, the suggested model was the best model in terms of accuracy. This research proposes evaluation criteria extracting method that combines topic extraction using LDA and refinement with k-nearest neighbor approach. This method overcomes the limits of previous dictionary-based models and frequency-based refinement models. This study can contribute to improve review analysis for deriving business insights in e-commerce market.

• Journal of Intelligence and Information Systems
• /
• v.26 no.1
• /
• pp.1-21
• /
• 2020

With the rapid development of artificial intelligence technology, various techniques have been developed to extract meaningful information from unstructured text data which constitutes a large portion of big data. Over the past decades, text mining technologies have been utilized in various industries for practical applications. In the field of business intelligence, it has been employed to discover new market and/or technology opportunities and support rational decision making of business participants. The market information such as market size, market growth rate, and market share is essential for setting companies' business strategies. There has been a continuous demand in various fields for specific product level-market information. However, the information has been generally provided at industry level or broad categories based on classification standards, making it difficult to obtain specific and proper information. In this regard, we propose a new methodology that can estimate the market sizes of product groups at more detailed levels than that of previously offered. We applied Word2Vec algorithm, a neural network based semantic word embedding model, to enable automatic market size estimation from individual companies' product information in a bottom-up manner. The overall process is as follows: First, the data related to product information is collected, refined, and restructured into suitable form for applying Word2Vec model. Next, the preprocessed data is embedded into vector space by Word2Vec and then the product groups are derived by extracting similar products names based on cosine similarity calculation. Finally, the sales data on the extracted products is summated to estimate the market size of the product groups. As an experimental data, text data of product names from Statistics Korea's microdata (345,103 cases) were mapped in multidimensional vector space by Word2Vec training. We performed parameters optimization for training and then applied vector dimension of 300 and window size of 15 as optimized parameters for further experiments. We employed index words of Korean Standard Industry Classification (KSIC) as a product name dataset to more efficiently cluster product groups. The product names which are similar to KSIC indexes were extracted based on cosine similarity. The market size of extracted products as one product category was calculated from individual companies' sales data. The market sizes of 11,654 specific product lines were automatically estimated by the proposed model. For the performance verification, the results were compared with actual market size of some items. The Pearson's correlation coefficient was 0.513. Our approach has several advantages differing from the previous studies. First, text mining and machine learning techniques were applied for the first time on market size estimation, overcoming the limitations of traditional sampling based- or multiple assumption required-methods. In addition, the level of market category can be easily and efficiently adjusted according to the purpose of information use by changing cosine similarity threshold. Furthermore, it has a high potential of practical applications since it can resolve unmet needs for detailed market size information in public and private sectors. Specifically, it can be utilized in technology evaluation and technology commercialization support program conducted by governmental institutions, as well as business strategies consulting and market analysis report publishing by private firms. The limitation of our study is that the presented model needs to be improved in terms of accuracy and reliability. The semantic-based word embedding module can be advanced by giving a proper order in the preprocessed dataset or by combining another algorithm such as Jaccard similarity with Word2Vec. Also, the methods of product group clustering can be changed to other types of unsupervised machine learning algorithm. Our group is currently working on subsequent studies and we expect that it can further improve the performance of the conceptually proposed basic model in this study.

• Nuclear Medicine and Molecular Imaging
• /
• v.40 no.5
• /
• pp.263-270
• /
• 2006

Purpose: Several radioisotope-labeled thymidine derivatives such as $[^{11}C]$thymidine was developed to demonstrate cell proliferation in tumor. But it is difficult to track metabolism with $[^{11}C]$thymidine due to rapid in vivo degradation and its short physical half-life. 3'-$[^{18}F]$fluoro-3'-deoxythymidine ($[^{18}F]$FLT) was reported to have the longer half life of fluorine-18 and the lack of metabolic degradation in vivo. Here, we described the synthesis of the 3'-$[^{18}F]$fluoro-3'-deoxythymidine ($[^{18}F]$FLT) and compared with $([^{18}F]FET)\;and\;([^{18}F]FDG)$ in cultured 9L cell and obtained the biodistribution and PET image in 9L tumor hearing rats. Material and Methods: For the synthesis of $[^{18}F]$FLT, 3-N-tert-butoxycarbonyl-(5'-O-(4,4'-dimet hoxytriphenylmethyl)-2'-deoxy-3'-O-(4-nitrobenzenesulfonyl)-${\beta}$-D-threopentofuranosyl)thymine was used as a FLT precursor, on which the tert-butyloxycarbonyl group was introduced to protect N3-position and nitrobenzenesulfonyl group. Radiolabeling of nosyl substitued precursor with $^{18}F$ was performed in acetonitrile at $120^{\circ}C$ and deproteced with 0.5 N HCI. The cell uptake was measured in cultured 9L glioma cell. The biodistribution was evaluated in 9L tumor bearing rats after intravenous injection at 10 min, 30 min, 60 min and 120 min and obtained PET image 60 minutes after injection. Results: The radiochemical yield was about 20-30% and radiochemical purity was more than 95% after HPLC purification. Cellular uptake of $[^{18}F]$FLT was increased as time elapsed. At 120 min post-injection, the ratios of tumor/blood, tumor/muscle and tumor/brain were $1.61{\pm}0.34,\;1.70{\pm}0.30\;and\;9.33{\pm}2.22$, respectively. The 9L tumor was well visualized at 60 min post injection in PET image. Conclusion: The uptake of $[^{18}F]$FLT in tumor was higher than in normal brain and PET image of $[^{18}F]$FLT was acceptable. These results suggest the possibility of $[^{18}F]$FLT at an imaging agent for brain tumor.

• Journal of the Korean Institute of Landscape Architecture
• /
• v.41 no.6
• /
• pp.20-28
• /
• 2013

The purpose of this study was to compare the ultra-violet(UV) radiation under the landscaping shade facilities and tree with natural solar UV of the outdoor space at summer middays. The UVA+B and UVB were recorded every minute from the $20^{th}$ of June to the $26^{th}$ of September 2012 at a height of 1.1m above in the four different shading conditions, with fours same measuring system consisting of two couple of analog UVA+B sensor(220~370nm, Genicom's GUVA-T21GH) and UVB sensor(220~320nm, Genicom's GUVA-T21GH) and data acquisition systems(Comfile Tech.'s Moacon). Four different shading conditions were under an wooden shelter($W4.2m{\times}L4.2m{\times}H2.5m$), a polyester membrane structure ($W4.9m{\times}L4.9m{\times}H2.6m$), a Salix koreensis($H11{\times}B30$), and a brick-paved plot without any shading material. Based on the 648 records of 17 sunny days, the time serial difference of natural solar UVA+B and UVB for midday periods were analysed and compared, and statistical analysis about the difference between the four shading conditions was done based on the 2,052 records of daytime period from 10 A.M. to 4 P.M.. The major findings were as follows; 1. The average UVA+B under the wooden shelter, the membrane and the tree were $39{\mu}W/cm^2$(3.4%), $74{\mu}W/cm^2$(6.4%), $87{\mu}W/cm^2$(7.6%) respectively, while the solar UVA+B was $1.148{\mu}W/cm^2$. Which means those facilities and tree screened at least 93% of solar UV+B. 2. The average UVB under the wooden shelter, the membrane and the tree were $12{\mu}W/cm^2$(5.8%), $26{\mu}W/cm^2$(13%), $17{\mu}W/cm^2$(8.2%) respectively, while the solar UVB was $207{\mu}W/cm^2$. The membrane showed the highest level and the wooden shelter lowest. 3. According to the results of time serial analysis, the difference between the three shaded conditions around noon was very small, but the differences of early morning and late afternoon were apparently big. Which seems caused by the matter of the formal and structural characteristics of the shading facilities and tree, not by the shading materials itself. In summary, the performance of the four landscaping shade facilities and tree were very good at screening the solar UV at outdoor of summer middays, but poor at screening the lateral UV during early morning and late afternoon. Therefore, it can be apparently said that the more delicate design of shading facilities and big tree or forest to block the additional lateral UV, the more effective in conditioning the outdoor space reducing the useless or even harmful radiation for human activities.

• Journal of Intelligence and Information Systems
• /
• v.18 no.3
• /
• pp.185-202
• /
• 2012

.Since the value of information has been realized in the information society, the usage and collection of information has become important. A facial expression that contains thousands of information as an artistic painting can be described in thousands of words. Followed by the idea, there has recently been a number of attempts to provide customers and companies with an intelligent service, which enables the perception of human emotions through one's facial expressions. For example, MIT Media Lab, the leading organization in this research area, has developed the human emotion prediction model, and has applied their studies to the commercial business. In the academic area, a number of the conventional methods such as Multiple Regression Analysis (MRA) or Artificial Neural Networks (ANN) have been applied to predict human emotion in prior studies. However, MRA is generally criticized because of its low prediction accuracy. This is inevitable since MRA can only explain the linear relationship between the dependent variables and the independent variable. To mitigate the limitations of MRA, some studies like Jung and Kim (2012) have used ANN as the alternative, and they reported that ANN generated more accurate prediction than the statistical methods like MRA. However, it has also been criticized due to over fitting and the difficulty of the network design (e.g. setting the number of the layers and the number of the nodes in the hidden layers). Under this background, we propose a novel model using Support Vector Regression (SVR) in order to increase the prediction accuracy. SVR is an extensive version of Support Vector Machine (SVM) designated to solve the regression problems. The model produced by SVR only depends on a subset of the training data, because the cost function for building the model ignores any training data that is close (within a threshold ${\varepsilon}$) to the model prediction. Using SVR, we tried to build a model that can measure the level of arousal and valence from the facial features. To validate the usefulness of the proposed model, we collected the data of facial reactions when providing appropriate visual stimulating contents, and extracted the features from the data. Next, the steps of the preprocessing were taken to choose statistically significant variables. In total, 297 cases were used for the experiment. As the comparative models, we also applied MRA and ANN to the same data set. For SVR, we adopted '${\varepsilon}$-insensitive loss function', and 'grid search' technique to find the optimal values of the parameters like C, d, ${\sigma}^2$, and ${\varepsilon}$. In the case of ANN, we adopted a standard three-layer backpropagation network, which has a single hidden layer. The learning rate and momentum rate of ANN were set to 10%, and we used sigmoid function as the transfer function of hidden and output nodes. We performed the experiments repeatedly by varying the number of nodes in the hidden layer to n/2, n, 3n/2, and 2n, where n is the number of the input variables. The stopping condition for ANN was set to 50,000 learning events. And, we used MAE (Mean Absolute Error) as the measure for performance comparison. From the experiment, we found that SVR achieved the highest prediction accuracy for the hold-out data set compared to MRA and ANN. Regardless of the target variables (the level of arousal, or the level of positive / negative valence), SVR showed the best performance for the hold-out data set. ANN also outperformed MRA, however, it showed the considerably lower prediction accuracy than SVR for both target variables. The findings of our research are expected to be useful to the researchers or practitioners who are willing to build the models for recognizing human emotions.

• Journal of Intelligence and Information Systems
• /
• v.18 no.3
• /
• pp.153-169
• /
• 2012

Exhibitions have played a key role of effective marketing activity which directly informs services and products to current and potential customers. Through participating in exhibitions, exhibitors have got the opportunity to make face-to-face contact so that they can secure the market share and improve their corporate images. According to this economic importance of exhibitions, show organizers try to adopt a new IT technology for improving their performance, and researchers have also studied services which can improve the satisfaction of visitors through analyzing visit patterns of visitors. Especially, as smart technologies make them monitor activities of visitors in real-time, they have considered booth recommender systems which infer preference of visitors and recommender proper service to them like on-line environment. However, while there are many studies which can improve their performance in the side of new technological development, they have not considered the choice factor of visitors for booth recommender systems. That is, studies for factors which can influence the development direction and effective diffusion of these systems are insufficient. Most of prior studies for the acceptance of new technologies and the continuous intention of use have adopted Technology Acceptance Model (TAM) and Extended Technology Acceptance Model (ETAM). Booth recommender systems may not be new technology because they are similar with commercial recommender systems such as book recommender systems, in the smart exhibition environment, they can be considered new technology. However, for considering the smart exhibition environment beyond TAM, measurements for the intention of reuse should focus on how booth recommender systems can provide correct information to visitors. In this study, through literature reviews, we draw factors which can influence the satisfaction and reuse intention of visitors for booth recommender systems, and design a model to forecast adaptation of visitors for booth recommendation in the exhibition environment. For these purposes, we conduct a survey for visitors who attended DMC Culture Open in November 2011 and experienced booth recommender systems using own smart phone, and examine hypothesis by regression analysis. As a result, factors which can influence the satisfaction of visitors for booth recommender systems are the effectiveness, perceived ease of use, argument quality, serendipity, and so on. Moreover, the satisfaction for booth recommender systems has a positive relationship with the development of reuse intention. For these results, we have some insights for booth recommender systems in the smart exhibition environment. First, this study gives shape to important factors which are considered when they establish strategies which induce visitors to consistently use booth recommender systems. Recently, although show organizers try to improve their performances using new IT technologies, their visitors have not felt the satisfaction from these efforts. At this point, this study can help them to provide services which can improve the satisfaction of visitors and make them last relationship with visitors. On the other hands, this study suggests that they managers along the using time of booth recommender systems. For example, in the early stage of the adoption, they should focus on the argument quality, perceived ease of use, and serendipity, so that improve the acceptance of booth recommender systems. After these stages, they should bridge the differences between expectation and perception for booth recommender systems, and lead continuous uses of visitors. However, this study has some limitations. We only use four factors which can influence the satisfaction of visitors. Therefore, we should development our model to consider important additional factors. And the exhibition in our experiments has small number of booths so that visitors may not need to booth recommender systems. In the future study, we will conduct experiments in the exhibition environment which has a larger scale.

• Journal of Intelligence and Information Systems
• /
• v.20 no.2
• /
• pp.73-92
• /
• 2014

An Adaptive Clustering-based Collaborative Filtering Technique was proposed to solve the fundamental problems of collaborative filtering, such as cold-start problems, scalability problems and data sparsity problems. Previous collaborative filtering techniques were carried out according to the recommendations based on the predicted preference of the user to a particular item using a similar item subset and a similar user subset composed based on the preference of users to items. For this reason, if the density of the user preference matrix is low, the reliability of the recommendation system will decrease rapidly. Therefore, the difficulty of creating a similar item subset and similar user subset will be increased. In addition, as the scale of service increases, the time needed to create a similar item subset and similar user subset increases geometrically, and the response time of the recommendation system is then increased. To solve these problems, this paper suggests a collaborative filtering technique that adapts a condition actively to the model and adopts the concepts of a context-based filtering technique. This technique consists of four major methodologies. First, items are made, the users are clustered according their feature vectors, and an inter-cluster preference between each item cluster and user cluster is then assumed. According to this method, the run-time for creating a similar item subset or user subset can be economized, the reliability of a recommendation system can be made higher than that using only the user preference information for creating a similar item subset or similar user subset, and the cold start problem can be partially solved. Second, recommendations are made using the prior composed item and user clusters and inter-cluster preference between each item cluster and user cluster. In this phase, a list of items is made for users by examining the item clusters in the order of the size of the inter-cluster preference of the user cluster, in which the user belongs, and selecting and ranking the items according to the predicted or recorded user preference information. Using this method, the creation of a recommendation model phase bears the highest load of the recommendation system, and it minimizes the load of the recommendation system in run-time. Therefore, the scalability problem and large scale recommendation system can be performed with collaborative filtering, which is highly reliable. Third, the missing user preference information is predicted using the item and user clusters. Using this method, the problem caused by the low density of the user preference matrix can be mitigated. Existing studies on this used an item-based prediction or user-based prediction. In this paper, Hao Ji's idea, which uses both an item-based prediction and user-based prediction, was improved. The reliability of the recommendation service can be improved by combining the predictive values of both techniques by applying the condition of the recommendation model. By predicting the user preference based on the item or user clusters, the time required to predict the user preference can be reduced, and missing user preference in run-time can be predicted. Fourth, the item and user feature vector can be made to learn the following input of the user feedback. This phase applied normalized user feedback to the item and user feature vector. This method can mitigate the problems caused by the use of the concepts of context-based filtering, such as the item and user feature vector based on the user profile and item properties. The problems with using the item and user feature vector are due to the limitation of quantifying the qualitative features of the items and users. Therefore, the elements of the user and item feature vectors are made to match one to one, and if user feedback to a particular item is obtained, it will be applied to the feature vector using the opposite one. Verification of this method was accomplished by comparing the performance with existing hybrid filtering techniques. Two methods were used for verification: MAE(Mean Absolute Error) and response time. Using MAE, this technique was confirmed to improve the reliability of the recommendation system. Using the response time, this technique was found to be suitable for a large scaled recommendation system. This paper suggested an Adaptive Clustering-based Collaborative Filtering Technique with high reliability and low time complexity, but it had some limitations. This technique focused on reducing the time complexity. Hence, an improvement in reliability was not expected. The next topic will be to improve this technique by rule-based filtering.

• Journal of Bio-Environment Control
• /
• v.28 no.1
• /
• pp.78-85
• /
• 2019

This study was designed to examine the performance of an aspirated radiation shield(ARS), which was made at the investigator's lab and characterized by relatively easier making and lower costs based on survey data and reports on errors in its measurements of temperature and relative humidity. The findings were summarized as follows: the ARS and the Jinju weather station made measurements and recorded the range of maximum, average, and minimum temperature at $2.0{\sim}34.1^{\circ}C$, $-6.1{\sim}22.2^{\circ}C$, $-14.0{\sim}15.1^{\circ}C$ and $0.4{\sim}31.5^{\circ}C$, $-5.8{\sim}22.0^{\circ}C$, $-14.1{\sim}16.3^{\circ}C$, respectively. There were no big differences in temperature measurements between the two institutions except that the lowest and highest point of maximum temperature was higher on the campus by $1.6^{\circ}C$ and $2.6^{\circ}C$, respectively. The measurements of ARS were tested against those of a standard thermometer. The results show that the temperature measured by ARS was lower by $-2.0^{\circ}C$ or higher by $1.8^{\circ}C$ than the temperature measured by a standard thermometer. The analysis results of its correlations with a standard thermometer reveal that the coefficient of determination was 0.99. Temperature was compared between fans and no fans, and the results show that maximum, average, and minimum temperature was higher overall with no fans by $0.5{\sim}7.6^{\circ}C$, $0.3{\sim}4.6^{\circ}C$ and $0.5{\sim}3.9^{\circ}C$, respectively. The daily average relative humidity measurements were compared between ARS and the weather station of Jinju, and the results show that the measurements of ARS were a little bit higher than those of the Jinju weather station. The measurements on June 27, July 26 and 29, and August 20 were relatively higher by 5.7%, 5.2%, 9.1%, and 5.8%, respectively, but differences in the monthly average between the two institutions were trivial at 2.0~3.0%. Relative humidity was in the range of -3.98~+7.78% overall based on measurements with ARS and Assman's psychometer. The study analyzed correlations in relative humidity between the measurements of the Jinju weather station and those of Assman's psychometer and found high correlations between them with the coefficient of determination at 0.94 and 0.97, respectively.

• Journal of Intelligence and Information Systems
• /
• v.25 no.1
• /
• pp.219-239
• /
• 2019

As the importance of providing customized services to individuals becomes important, researches on personalized recommendation systems are constantly being carried out. Collaborative filtering is one of the most popular systems in academia and industry. However, there exists limitation in a sense that recommendations were mostly based on quantitative information such as users' ratings, which made the accuracy be lowered. To solve these problems, many studies have been actively attempted to improve the performance of the recommendation system by using other information besides the quantitative information. Good examples are the usages of the sentiment analysis on customer review text data. Nevertheless, the existing research has not directly combined the results of the sentiment analysis and quantitative rating scores in the recommendation system. Therefore, this study aims to reflect the sentiments shown in the reviews into the rating scores. In other words, we propose a new algorithm that can directly convert the user 's own review into the empirically quantitative information and reflect it directly to the recommendation system. To do this, we needed to quantify users' reviews, which were originally qualitative information. In this study, sentiment score was calculated through sentiment analysis technique of text mining. The data was targeted for movie review. Based on the data, a domain specific sentiment dictionary is constructed for the movie reviews. Regression analysis was used as a method to construct sentiment dictionary. Each positive / negative dictionary was constructed using Lasso regression, Ridge regression, and ElasticNet methods. Based on this constructed sentiment dictionary, the accuracy was verified through confusion matrix. The accuracy of the Lasso based dictionary was 70%, the accuracy of the Ridge based dictionary was 79%, and that of the ElasticNet (${\alpha}=0.3$) was 83%. Therefore, in this study, the sentiment score of the review is calculated based on the dictionary of the ElasticNet method. It was combined with a rating to create a new rating. In this paper, we show that the collaborative filtering that reflects sentiment scores of user review is superior to the traditional method that only considers the existing rating. In order to show that the proposed algorithm is based on memory-based user collaboration filtering, item-based collaborative filtering and model based matrix factorization SVD, and SVD ++. Based on the above algorithm, the mean absolute error (MAE) and the root mean square error (RMSE) are calculated to evaluate the recommendation system with a score that combines sentiment scores with a system that only considers scores. When the evaluation index was MAE, it was improved by 0.059 for UBCF, 0.0862 for IBCF, 0.1012 for SVD and 0.188 for SVD ++. When the evaluation index is RMSE, UBCF is 0.0431, IBCF is 0.0882, SVD is 0.1103, and SVD ++ is 0.1756. As a result, it can be seen that the prediction performance of the evaluation point reflecting the sentiment score proposed in this paper is superior to that of the conventional evaluation method. In other words, in this paper, it is confirmed that the collaborative filtering that reflects the sentiment score of the user review shows superior accuracy as compared with the conventional type of collaborative filtering that only considers the quantitative score. We then attempted paired t-test validation to ensure that the proposed model was a better approach and concluded that the proposed model is better. In this study, to overcome limitations of previous researches that judge user's sentiment only by quantitative rating score, the review was numerically calculated and a user's opinion was more refined and considered into the recommendation system to improve the accuracy. The findings of this study have managerial implications to recommendation system developers who need to consider both quantitative information and qualitative information it is expect. The way of constructing the combined system in this paper might be directly used by the developers.

• Journal of Intelligence and Information Systems
• /
• v.27 no.3
• /
• pp.139-156
• /
• 2021

The biggest reason for using a deep learning model in image classification is that it is possible to consider the relationship between each region by extracting each region's features from the overall information of the image. However, the CNN model may not be suitable for emotional image data without the image's regional features. To solve the difficulty of classifying emotion images, many researchers each year propose a CNN-based architecture suitable for emotion images. Studies on the relationship between color and human emotion were also conducted, and results were derived that different emotions are induced according to color. In studies using deep learning, there have been studies that apply color information to image subtraction classification. The case where the image's color information is additionally used than the case where the classification model is trained with only the image improves the accuracy of classifying image emotions. This study proposes two ways to increase the accuracy by incorporating the result value after the model classifies an image's emotion. Both methods improve accuracy by modifying the result value based on statistics using the color of the picture. When performing the test by finding the two-color combinations most distributed for all training data, the two-color combinations most distributed for each test data image were found. The result values were corrected according to the color combination distribution. This method weights the result value obtained after the model classifies an image's emotion by creating an expression based on the log function and the exponential function. Emotion6, classified into six emotions, and Artphoto classified into eight categories were used for the image data. Densenet169, Mnasnet, Resnet101, Resnet152, and Vgg19 architectures were used for the CNN model, and the performance evaluation was compared before and after applying the two-stage learning to the CNN model. Inspired by color psychology, which deals with the relationship between colors and emotions, when creating a model that classifies an image's sentiment, we studied how to improve accuracy by modifying the result values based on color. Sixteen colors were used: red, orange, yellow, green, blue, indigo, purple, turquoise, pink, magenta, brown, gray, silver, gold, white, and black. It has meaning. Using Scikit-learn's Clustering, the seven colors that are primarily distributed in the image are checked. Then, the RGB coordinate values of the colors from the image are compared with the RGB coordinate values of the 16 colors presented in the above data. That is, it was converted to the closest color. Suppose three or more color combinations are selected. In that case, too many color combinations occur, resulting in a problem in which the distribution is scattered, so a situation fewer influences the result value. Therefore, to solve this problem, two-color combinations were found and weighted to the model. Before training, the most distributed color combinations were found for all training data images. The distribution of color combinations for each class was stored in a Python dictionary format to be used during testing. During the test, the two-color combinations that are most distributed for each test data image are found. After that, we checked how the color combinations were distributed in the training data and corrected the result. We devised several equations to weight the result value from the model based on the extracted color as described above. The data set was randomly divided by 80:20, and the model was verified using 20% of the data as a test set. After splitting the remaining 80% of the data into five divisions to perform 5-fold cross-validation, the model was trained five times using different verification datasets. Finally, the performance was checked using the test dataset that was previously separated. Adam was used as the activation function, and the learning rate was set to 0.01. The training was performed as much as 20 epochs, and if the validation loss value did not decrease during five epochs of learning, the experiment was stopped. Early tapping was set to load the model with the best validation loss value. The classification accuracy was better when the extracted information using color properties was used together than the case using only the CNN architecture.