• Journal of Intelligence and Information Systems
• /
• v.20 no.3
• /
• pp.93-108
• /
• 2014

To support business decision making, interests and efforts to analyze and use transaction data in different perspectives are increasing. Such efforts are not only limited to customer management or marketing, but also used for monitoring and detecting fraud transactions. Fraud transactions are evolving into various patterns by taking advantage of information technology. To reflect the evolution of fraud transactions, there are many efforts on fraud detection methods and advanced application systems in order to improve the accuracy and ease of fraud detection. As a case of fraud detection, this study aims to provide effective fraud detection methods for auction exception agricultural products in the largest Korean agricultural wholesale market. Auction exception products policy exists to complement auction-based trades in agricultural wholesale market. That is, most trades on agricultural products are performed by auction; however, specific products are assigned as auction exception products when total volumes of products are relatively small, the number of wholesalers is small, or there are difficulties for wholesalers to purchase the products. However, auction exception products policy makes several problems on fairness and transparency of transaction, which requires help of fraud detection. In this study, to generate fraud detection rules, real huge agricultural products trade transaction data from 2008 to 2010 in the market are analyzed, which increase more than 1 million transactions and 1 billion US dollar in transaction volume. Agricultural transaction data has unique characteristics such as frequent changes in supply volumes and turbulent time-dependent changes in price. Since this was the first trial to identify fraud transactions in this domain, there was no training data set for supervised learning. So, fraud detection rules are generated using outlier detection approach. We assume that outlier transactions have more possibility of fraud transactions than normal transactions. The outlier transactions are identified to compare daily average unit price, weekly average unit price, and quarterly average unit price of product items. Also quarterly averages unit price of product items of the specific wholesalers are used to identify outlier transactions. The reliability of generated fraud detection rules are confirmed by domain experts. To determine whether a transaction is fraudulent or not, normal distribution and normalized Z-value concept are applied. That is, a unit price of a transaction is transformed to Z-value to calculate the occurrence probability when we approximate the distribution of unit prices to normal distribution. The modified Z-value of the unit price in the transaction is used rather than using the original Z-value of it. The reason is that in the case of auction exception agricultural products, Z-values are influenced by outlier fraud transactions themselves because the number of wholesalers is small. The modified Z-values are called Self-Eliminated Z-scores because they are calculated excluding the unit price of the specific transaction which is subject to check whether it is fraud transaction or not. To show the usefulness of the proposed approach, a prototype of fraud transaction detection system is developed using Delphi. The system consists of five main menus and related submenus. First functionalities of the system is to import transaction databases. Next important functions are to set up fraud detection parameters. By changing fraud detection parameters, system users can control the number of potential fraud transactions. Execution functions provide fraud detection results which are found based on fraud detection parameters. The potential fraud transactions can be viewed on screen or exported as files. The study is an initial trial to identify fraud transactions in Auction Exception Agricultural Products. There are still many remained research topics of the issue. First, the scope of analysis data was limited due to the availability of data. It is necessary to include more data on transactions, wholesalers, and producers to detect fraud transactions more accurately. Next, we need to extend the scope of fraud transaction detection to fishery products. Also there are many possibilities to apply different data mining techniques for fraud detection. For example, time series approach is a potential technique to apply the problem. Even though outlier transactions are detected based on unit prices of transactions, however it is possible to derive fraud detection rules based on transaction volumes.

• Journal of Intelligence and Information Systems
• /
• v.23 no.2
• /
• pp.19-38
• /
• 2017

Collaborative filtering (CF) algorithm has been popularly used for implementing recommender systems. Until now, there have been many prior studies to improve the accuracy of CF. Among them, some recent studies adopt 'hybrid recommendation approach', which enhances the performance of conventional CF by using additional information. In this research, we propose a new hybrid recommender system which fuses CF and the results from the social network analysis on trust and distrust relationship networks among users to enhance prediction accuracy. The proposed algorithm of our study is based on memory-based CF. But, when calculating the similarity between users in CF, our proposed algorithm considers not only the correlation of the users' numeric rating patterns, but also the users' in-degree centrality values derived from trust and distrust relationship networks. In specific, it is designed to amplify the similarity between a target user and his or her neighbor when the neighbor has higher in-degree centrality in the trust relationship network. Also, it attenuates the similarity between a target user and his or her neighbor when the neighbor has higher in-degree centrality in the distrust relationship network. Our proposed algorithm considers four (4) types of user relationships - direct trust, indirect trust, direct distrust, and indirect distrust - in total. And, it uses four adjusting coefficients, which adjusts the level of amplification / attenuation for in-degree centrality values derived from direct / indirect trust and distrust relationship networks. To determine optimal adjusting coefficients, genetic algorithms (GA) has been adopted. Under this background, we named our proposed algorithm as SNACF-GA (Social Network Analysis - based CF using GA). To validate the performance of the SNACF-GA, we used a real-world data set which is called 'Extended Epinions dataset' provided by 'trustlet.org'. It is the data set contains user responses (rating scores and reviews) after purchasing specific items (e.g. car, movie, music, book) as well as trust / distrust relationship information indicating whom to trust or distrust between users. The experimental system was basically developed using Microsoft Visual Basic for Applications (VBA), but we also used UCINET 6 for calculating the in-degree centrality of trust / distrust relationship networks. In addition, we used Palisade Software's Evolver, which is a commercial software implements genetic algorithm. To examine the effectiveness of our proposed system more precisely, we adopted two comparison models. The first comparison model is conventional CF. It only uses users' explicit numeric ratings when calculating the similarities between users. That is, it does not consider trust / distrust relationship between users at all. The second comparison model is SNACF (Social Network Analysis - based CF). SNACF differs from the proposed algorithm SNACF-GA in that it considers only direct trust / distrust relationships. It also does not use GA optimization. The performances of the proposed algorithm and comparison models were evaluated by using average MAE (mean absolute error). Experimental result showed that the optimal adjusting coefficients for direct trust, indirect trust, direct distrust, indirect distrust were 0, 1.4287, 1.5, 0.4615 each. This implies that distrust relationships between users are more important than trust ones in recommender systems. From the perspective of recommendation accuracy, SNACF-GA (Avg. MAE = 0.111943), the proposed algorithm which reflects both direct and indirect trust / distrust relationships information, was found to greatly outperform a conventional CF (Avg. MAE = 0.112638). Also, the algorithm showed better recommendation accuracy than the SNACF (Avg. MAE = 0.112209). To confirm whether these differences are statistically significant or not, we applied paired samples t-test. The results from the paired samples t-test presented that the difference between SNACF-GA and conventional CF was statistical significant at the 1% significance level, and the difference between SNACF-GA and SNACF was statistical significant at the 5%. Our study found that the trust/distrust relationship can be important information for improving performance of recommendation algorithms. Especially, distrust relationship information was found to have a greater impact on the performance improvement of CF. This implies that we need to have more attention on distrust (negative) relationships rather than trust (positive) ones when tracking and managing social relationships between users.

• Journal of Intelligence and Information Systems
• /
• v.26 no.3
• /
• pp.1-36
• /
• 2020

This study is to reveal the acceptance factors of the Market Sentiment Index (MSI) created by reflecting the investor sentiment extracted by processing unstructured big data. The research model was established by exploring exogenous variables based on the rational behavior theory and applying the Technology Acceptance Model (TAM). The acceptance of MSI provided to investors in the stock market was found to be influenced by the exogenous variables presented in this study. The results of causal analysis are as follows. First, self-efficacy, investment opportunities, Innovativeness, and perceived cost significantly affect perceived ease of use. Second, Diversity of services and perceived benefits have a statistically significant impact on perceived usefulness. Third, Perceived ease of use and perceived usefulness have a statistically significant effect on attitude to use. Fourth, Attitude to use statistically significantly influences the intention to use, and the investment opportunities as an independent variable affects the intention to use. Fifth, the intention to use statistically significantly affects the final dependent variable, the intention to use continuously. The mediating effect between the independent and dependent variables of the research model is as follows. First, The indirect effect on the causal route from diversity of services to continuous use intention was 0.1491, which was statistically significant at the significance level of 1%. Second, The indirect effect on the causal route from perceived benefit to continuous use intention was 0.1281, which was statistically significant at the significance level of 1%. The results of the multi-group analysis are as follows. First, for groups with and without stock investment experience, multi-group analysis was not possible because the measurement uniformity between the two groups was not secured. Second, the analysis result of the difference in the effect of independent variables of male and female groups on the intention to use continuously, where measurement uniformity was secured between the two groups, In the causal route from usage attitude to usage intention, women are higher than men. And in the causal route from use intention to continuous use intention, males were very high and showed statistically significant difference at significance level 5%.

• Journal of Intelligence and Information Systems
• /
• v.18 no.1
• /
• pp.23-38
• /
• 2012

There is a large difference between purchasing patterns in an online shopping mall and in an offline market. This difference may be caused mainly by the difference in accessibility of online and offline markets. It means that an interval between the initial purchasing decision and its realization appears to be relatively short in an online shopping mall, because a customer can make an order immediately. Because of the short interval between a purchasing decision and its realization, an online shopping mall transaction usually contains fewer items than that of an offline market. In an offline market, customers usually keep some items in mind and buy them all at once a few days after deciding to buy them, instead of buying each item individually and immediately. On the contrary, more than 70% of online shopping mall transactions contain only one item. This statistic implies that traditional data mining techniques cannot be directly applied to online market analysis, because hardly any association rules can survive with an acceptable level of Support because of too many Null Transactions. Most market basket analyses on online shopping mall transactions, therefore, have been performed by expanding the co-occurrence criteria of traditional association rule mining. While the traditional co-occurrence criteria defines items purchased in one transaction as concurrently purchased items, the expanded co-occurrence criteria regards items purchased by a customer during some predefined period (e.g., a day) as concurrently purchased items. In studies using expanded co-occurrence criteria, however, the criteria has been defined arbitrarily by researchers without any theoretical grounds or agreement. The lack of clear grounds of adopting a certain co-occurrence criteria degrades the reliability of the analytical results. Moreover, it is hard to derive new meaningful findings by combining the outcomes of previous individual studies. In this paper, we attempt to compare expanded co-occurrence criteria and propose a guideline for selecting an appropriate one. First of all, we compare the accuracy of association rules discovered according to various co-occurrence criteria. By doing this experiment we expect that we can provide a guideline for selecting appropriate co-occurrence criteria that corresponds to the purpose of the analysis. Additionally, we will perform similar experiments with several groups of customers that are segmented by each customer's average duration between orders. By this experiment, we attempt to discover the relationship between the optimal co-occurrence criteria and the customer's average duration between orders. Finally, by a series of experiments, we expect that we can provide basic guidelines for developing customized recommendation systems. Our experiments use a real dataset acquired from one of the largest internet shopping malls in Korea. We use 66,278 transactions of 3,847 customers conducted during the last two years. Overall results show that the accuracy of association rules of frequent shoppers (whose average duration between orders is relatively short) is higher than that of causal shoppers. In addition we discover that with frequent shoppers, the accuracy of association rules appears very high when the co-occurrence criteria of the training set corresponds to the validation set (i.e., target set). It implies that the co-occurrence criteria of frequent shoppers should be set according to the application purpose period. For example, an analyzer should use a day as a co-occurrence criterion if he/she wants to offer a coupon valid only for a day to potential customers who will use the coupon. On the contrary, an analyzer should use a month as a co-occurrence criterion if he/she wants to publish a coupon book that can be used for a month. In the case of causal shoppers, the accuracy of association rules appears to not be affected by the period of the application purposes. The accuracy of the causal shoppers' association rules becomes higher when the longer co-occurrence criterion has been adopted. It implies that an analyzer has to set the co-occurrence criterion for as long as possible, regardless of the application purpose period.

• Journal of Intelligence and Information Systems
• /
• v.28 no.3
• /
• pp.237-258
• /
• 2022

As the global market environment changes, the domestic ICT industry has a growing influence on the world economy. This industry is regarded as an important driving force in the national economy from a technological and social point of view. In particular, small and medium-sized enterprises (SMEs) in the ICT industry are regarded as essential actors of domestic economic development in terms of company diversity, technology development and job creation. However, since it is small compared to large-sized enterprises, it is difficult for SMEs to survive with a differentiated strategy in an incomplete and rapidly changing environment. Therefore, SMEs must make a lot of efforts to improve their own capabilities, and the government needs to provide the desirable help suitable for corporate internal resources so that they can continue to be competitive. This study classifies the types of ICT SMEs participating in government support programs, and analyzes the relationship between resources and performance creation of each type. The data from the "ICT Small and Medium Enterprises Survey" conducted annually by the Ministry of Science and ICT was used. In the first stage, ICT SMEs were clustered based on common factors according to their experiences with government support programs. Three clusters were meaningfully classified, and each cluster was named "active participation type," "initial support type," and "soloist type." As a second step, this study compared the characteristics of each cluster through profiling analysis for each cluster. The third step carried out in this study was to find out the mechanism of R&D performance creation for each cluster through regression analysis. Different factors affected performance creation for each cluster, and the magnitude of the influence was also different. Specifically, for "active participation type", "current manpower", "technology competitiveness", and "R&D investment in the previous year" were found to be important factors in creating R&D performance. "Initial support type" was identified as "whether or not a dedicated R&D organization exists", "R&D investment amount in the previous year", "Ratio of sales to large companies", and "Ratio of vendors supplied to large companies" contributed to the performance. Lastly, in the case of "soloist type", "current workforce" and "future recruitment plan", "technological competitiveness", "R&D investment", "large company sales ratio", and "overseas sales ratio" showed a significant relationship with the performance. This study has practical implications of showing what strategy should be established when supporting SMEs in the future according to the government's participation in the startup program and providing a guide on what kind of support should be provided.

• Journal of Intelligence and Information Systems
• /
• v.17 no.4
• /
• pp.227-240
• /
• 2011

New concepts and ideas often result from extensive recombination of existing concepts or ideas. Both researchers and developers build on existing concepts and ideas in published papers or registered patents to develop new theories and technologies that in turn serve as a basis for further development. As the importance of patent increases, so does that of patent analysis. Patent analysis is largely divided into network-based and keyword-based analyses. The former lacks its ability to analyze information technology in details while the letter is unable to identify the relationship between such technologies. In order to overcome the limitations of network-based and keyword-based analyses, this study, which blends those two methods, suggests the keyword network based analysis methodology. In this study, we collected significant technology information in each patent that is related to Light Emitting Diode (LED) through text mining, built a keyword network, and then executed a community network analysis on the collected data. The results of analysis are as the following. First, the patent keyword network indicated very low density and exceptionally high clustering coefficient. Technically, density is obtained by dividing the number of ties in a network by the number of all possible ties. The value ranges between 0 and 1, with higher values indicating denser networks and lower values indicating sparser networks. In real-world networks, the density varies depending on the size of a network; increasing the size of a network generally leads to a decrease in the density. The clustering coefficient is a network-level measure that illustrates the tendency of nodes to cluster in densely interconnected modules. This measure is to show the small-world property in which a network can be highly clustered even though it has a small average distance between nodes in spite of the large number of nodes. Therefore, high density in patent keyword network means that nodes in the patent keyword network are connected sporadically, and high clustering coefficient shows that nodes in the network are closely connected one another. Second, the cumulative degree distribution of the patent keyword network, as any other knowledge network like citation network or collaboration network, followed a clear power-law distribution. A well-known mechanism of this pattern is the preferential attachment mechanism, whereby a node with more links is likely to attain further new links in the evolution of the corresponding network. Unlike general normal distributions, the power-law distribution does not have a representative scale. This means that one cannot pick a representative or an average because there is always a considerable probability of finding much larger values. Networks with power-law distributions are therefore often referred to as scale-free networks. The presence of heavy-tailed scale-free distribution represents the fundamental signature of an emergent collective behavior of the actors who contribute to forming the network. In our context, the more frequently a patent keyword is used, the more often it is selected by researchers and is associated with other keywords or concepts to constitute and convey new patents or technologies. The evidence of power-law distribution implies that the preferential attachment mechanism suggests the origin of heavy-tailed distributions in a wide range of growing patent keyword network. Third, we found that among keywords that flew into a particular field, the vast majority of keywords with new links join existing keywords in the associated community in forming the concept of a new patent. This finding resulted in the same outcomes for both the short-term period (4-year) and long-term period (10-year) analyses. Furthermore, using the keyword combination information that was derived from the methodology suggested by our study enables one to forecast which concepts combine to form a new patent dimension and refer to those concepts when developing a new patent.

• Journal of Intelligence and Information Systems
• /
• v.18 no.4
• /
• pp.117-127
• /
• 2012

Due to the recent expansion of the Web 2.0 -based services, along with the widespread of smartphones, online social network services are being popularized among users. Online social network services are the online community services which enable users to communicate each other, share information and expand human relationships. In the social network services, each relation between users is represented by a graph consisting of nodes and links. As the users of online social network services are increasing rapidly, the SNS are actively utilized in enterprise marketing, analysis of social phenomenon and so on. Social Network Analysis (SNA) is the systematic way to analyze social relationships among the members of the social network using the network theory. In general social network theory consists of nodes and arcs, and it is often depicted in a social network diagram. In a social network diagram, nodes represent individual actors within the network and arcs represent relationships between the nodes. With SNA, we can measure relationships among the people such as degree of intimacy, intensity of connection and classification of the groups. Ever since Social Networking Services (SNS) have drawn increasing attention from millions of users, numerous researches have made to analyze their user relationships and messages. There are typical representative SNA methods: degree centrality, betweenness centrality and closeness centrality. In the degree of centrality analysis, the shortest path between nodes is not considered. However, it is used as a crucial factor in betweenness centrality, closeness centrality and other SNA methods. In previous researches in SNA, the computation time was not too expensive since the size of social network was small. Unfortunately, most SNA methods require significant time to process relevant data, and it makes difficult to apply the ever increasing SNS data in social network studies. For instance, if the number of nodes in online social network is n, the maximum number of link in social network is n(n-1)/2. It means that it is too expensive to analyze the social network, for example, if the number of nodes is 10,000 the number of links is 49,995,000. Therefore, we propose a heuristic-based method for finding the shortest path among users in the SNS user graph. Through the shortest path finding method, we will show how efficient our proposed approach may be by conducting betweenness centrality analysis and closeness centrality analysis, both of which are widely used in social network studies. Moreover, we devised an enhanced method with addition of best-first-search method and preprocessing step for the reduction of computation time and rapid search of the shortest paths in a huge size of online social network. Best-first-search method finds the shortest path heuristically, which generalizes human experiences. As large number of links is shared by only a few nodes in online social networks, most nods have relatively few connections. As a result, a node with multiple connections functions as a hub node. When searching for a particular node, looking for users with numerous links instead of searching all users indiscriminately has a better chance of finding the desired node more quickly. In this paper, we employ the degree of user node vn as heuristic evaluation function in a graph G = (N, E), where N is a set of vertices, and E is a set of links between two different nodes. As the heuristic evaluation function is used, the worst case could happen when the target node is situated in the bottom of skewed tree. In order to remove such a target node, the preprocessing step is conducted. Next, we find the shortest path between two nodes in social network efficiently and then analyze the social network. For the verification of the proposed method, we crawled 160,000 people from online and then constructed social network. Then we compared with previous methods, which are best-first-search and breath-first-search, in time for searching and analyzing. The suggested method takes 240 seconds to search nodes where breath-first-search based method takes 1,781 seconds (7.4 times faster). Moreover, for social network analysis, the suggested method is 6.8 times and 1.8 times faster than betweenness centrality analysis and closeness centrality analysis, respectively. The proposed method in this paper shows the possibility to analyze a large size of social network with the better performance in time. As a result, our method would improve the efficiency of social network analysis, making it particularly useful in studying social trends or phenomena.

• Journal of Intelligence and Information Systems
• /
• v.26 no.4
• /
• pp.127-148
• /
• 2020

The data center is a physical environment facility for accommodating computer systems and related components, and is an essential foundation technology for next-generation core industries such as big data, smart factories, wearables, and smart homes. In particular, with the growth of cloud computing, the proportional expansion of the data center infrastructure is inevitable. Monitoring the health of these data center facilities is a way to maintain and manage the system and prevent failure. If a failure occurs in some elements of the facility, it may affect not only the relevant equipment but also other connected equipment, and may cause enormous damage. In particular, IT facilities are irregular due to interdependence and it is difficult to know the cause. In the previous study predicting failure in data center, failure was predicted by looking at a single server as a single state without assuming that the devices were mixed. Therefore, in this study, data center failures were classified into failures occurring inside the server (Outage A) and failures occurring outside the server (Outage B), and focused on analyzing complex failures occurring within the server. Server external failures include power, cooling, user errors, etc. Since such failures can be prevented in the early stages of data center facility construction, various solutions are being developed. On the other hand, the cause of the failure occurring in the server is difficult to determine, and adequate prevention has not yet been achieved. In particular, this is the reason why server failures do not occur singularly, cause other server failures, or receive something that causes failures from other servers. In other words, while the existing studies assumed that it was a single server that did not affect the servers and analyzed the failure, in this study, the failure occurred on the assumption that it had an effect between servers. In order to define the complex failure situation in the data center, failure history data for each equipment existing in the data center was used. There are four major failures considered in this study: Network Node Down, Server Down, Windows Activation Services Down, and Database Management System Service Down. The failures that occur for each device are sorted in chronological order, and when a failure occurs in a specific equipment, if a failure occurs in a specific equipment within 5 minutes from the time of occurrence, it is defined that the failure occurs simultaneously. After configuring the sequence for the devices that have failed at the same time, 5 devices that frequently occur simultaneously within the configured sequence were selected, and the case where the selected devices failed at the same time was confirmed through visualization. Since the server resource information collected for failure analysis is in units of time series and has flow, we used Long Short-term Memory (LSTM), a deep learning algorithm that can predict the next state through the previous state. In addition, unlike a single server, the Hierarchical Attention Network deep learning model structure was used in consideration of the fact that the level of multiple failures for each server is different. This algorithm is a method of increasing the prediction accuracy by giving weight to the server as the impact on the failure increases. The study began with defining the type of failure and selecting the analysis target. In the first experiment, the same collected data was assumed as a single server state and a multiple server state, and compared and analyzed. The second experiment improved the prediction accuracy in the case of a complex server by optimizing each server threshold. In the first experiment, which assumed each of a single server and multiple servers, in the case of a single server, it was predicted that three of the five servers did not have a failure even though the actual failure occurred. However, assuming multiple servers, all five servers were predicted to have failed. As a result of the experiment, the hypothesis that there is an effect between servers is proven. As a result of this study, it was confirmed that the prediction performance was superior when the multiple servers were assumed than when the single server was assumed. In particular, applying the Hierarchical Attention Network algorithm, assuming that the effects of each server will be different, played a role in improving the analysis effect. In addition, by applying a different threshold for each server, the prediction accuracy could be improved. This study showed that failures that are difficult to determine the cause can be predicted through historical data, and a model that can predict failures occurring in servers in data centers is presented. It is expected that the occurrence of disability can be prevented in advance using the results of this study.

• Journal of Intelligence and Information Systems
• /
• v.19 no.4
• /
• pp.123-132
• /
• 2013

As the smartphones are equipped with various sensors such as the accelerometer, GPS, gravity sensor, gyros, ambient light sensor, proximity sensor, and so on, there have been many research works on making use of these sensors to create valuable applications. Human activity recognition is one such application that is motivated by various welfare applications such as the support for the elderly, measurement of calorie consumption, analysis of lifestyles, analysis of exercise patterns, and so on. One of the challenges faced when using the smartphone sensors for activity recognition is that the number of sensors used should be minimized to save the battery power. When the number of sensors used are restricted, it is difficult to realize a highly accurate activity recognizer or a classifier because it is hard to distinguish between subtly different activities relying on only limited information. The difficulty gets especially severe when the number of different activity classes to be distinguished is very large. In this paper, we show that a fairly accurate classifier can be built that can distinguish ten different activities by using only a single sensor data, i.e., the smartphone accelerometer data. The approach that we take to dealing with this ten-class problem is to use the ensemble of nested dichotomy (END) method that transforms a multi-class problem into multiple two-class problems. END builds a committee of binary classifiers in a nested fashion using a binary tree. At the root of the binary tree, the set of all the classes are split into two subsets of classes by using a binary classifier. At a child node of the tree, a subset of classes is again split into two smaller subsets by using another binary classifier. Continuing in this way, we can obtain a binary tree where each leaf node contains a single class. This binary tree can be viewed as a nested dichotomy that can make multi-class predictions. Depending on how a set of classes are split into two subsets at each node, the final tree that we obtain can be different. Since there can be some classes that are correlated, a particular tree may perform better than the others. However, we can hardly identify the best tree without deep domain knowledge. The END method copes with this problem by building multiple dichotomy trees randomly during learning, and then combining the predictions made by each tree during classification. The END method is generally known to perform well even when the base learner is unable to model complex decision boundaries As the base classifier at each node of the dichotomy, we have used another ensemble classifier called the random forest. A random forest is built by repeatedly generating a decision tree each time with a different random subset of features using a bootstrap sample. By combining bagging with random feature subset selection, a random forest enjoys the advantage of having more diverse ensemble members than a simple bagging. As an overall result, our ensemble of nested dichotomy can actually be seen as a committee of committees of decision trees that can deal with a multi-class problem with high accuracy. The ten classes of activities that we distinguish in this paper are 'Sitting', 'Standing', 'Walking', 'Running', 'Walking Uphill', 'Walking Downhill', 'Running Uphill', 'Running Downhill', 'Falling', and 'Hobbling'. The features used for classifying these activities include not only the magnitude of acceleration vector at each time point but also the maximum, the minimum, and the standard deviation of vector magnitude within a time window of the last 2 seconds, etc. For experiments to compare the performance of END with those of other methods, the accelerometer data has been collected at every 0.1 second for 2 minutes for each activity from 5 volunteers. Among these 5,900 ($=5{\times}(60{\times}2-2)/0.1$) data collected for each activity (the data for the first 2 seconds are trashed because they do not have time window data), 4,700 have been used for training and the rest for testing. Although 'Walking Uphill' is often confused with some other similar activities, END has been found to classify all of the ten activities with a fairly high accuracy of 98.4%. On the other hand, the accuracies achieved by a decision tree, a k-nearest neighbor, and a one-versus-rest support vector machine have been observed as 97.6%, 96.5%, and 97.6%, respectively.

• 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.