• Journal of Intelligence and Information Systems
• /
• v.28 no.2
• /
• pp.191-206
• /
• 2022

Recently, it is a de-facto approach to utilize a pre-trained language model(PLM) to achieve the state-of-the-art performance for various natural language tasks(called downstream tasks) such as sentiment analysis and question answering. However, similar to any other machine learning method, PLM tends to depend on the data distribution seen during the training phase and shows worse performance on the unseen (Out-of-Distribution) domain. Due to the aforementioned reason, there have been many efforts to develop domain-specified PLM for various fields such as medical and legal industries. In this paper, we discuss the training of a finance domain-specified PLM for the Korean language and its applications. Our finance domain-specified PLM, KB-BERT, is trained on a carefully curated financial corpus that includes domain-specific documents such as financial reports. We provide extensive performance evaluation results on three natural language tasks, topic classification, sentiment analysis, and question answering. Compared to the state-of-the-art Korean PLM models such as KoELECTRA and KLUE-RoBERTa, KB-BERT shows comparable performance on general datasets based on common corpora like Wikipedia and news articles. Moreover, KB-BERT outperforms compared models on finance domain datasets that require finance-specific knowledge to solve given problems.

• Journal of Intelligence and Information Systems
• /
• v.26 no.4
• /
• pp.111-126
• /
• 2020

Most industries have recently become aware of the importance of customer lifetime value as they are exposed to a competitive environment. As a result, preventing customers from churn is becoming a more important business issue than securing new customers. This is because maintaining churn customers is far more economical than securing new customers, and in fact, the acquisition cost of new customers is known to be five to six times higher than the maintenance cost of churn customers. Also, Companies that effectively prevent customer churn and improve customer retention rates are known to have a positive effect on not only increasing the company's profitability but also improving its brand image by improving customer satisfaction. Predicting customer churn, which had been conducted as a sub-research area for CRM, has recently become more important as a big data-based performance marketing theme due to the development of business machine learning technology. Until now, research on customer churn prediction has been carried out actively in such sectors as the mobile telecommunication industry, the financial industry, the distribution industry, and the game industry, which are highly competitive and urgent to manage churn. In addition, These churn prediction studies were focused on improving the performance of the churn prediction model itself, such as simply comparing the performance of various models, exploring features that are effective in forecasting departures, or developing new ensemble techniques, and were limited in terms of practical utilization because most studies considered the entire customer group as a group and developed a predictive model. As such, the main purpose of the existing related research was to improve the performance of the predictive model itself, and there was a relatively lack of research to improve the overall customer churn prediction process. In fact, customers in the business have different behavior characteristics due to heterogeneous transaction patterns, and the resulting churn rate is different, so it is unreasonable to assume the entire customer as a single customer group. Therefore, it is desirable to segment customers according to customer classification criteria, such as loyalty, and to operate an appropriate churn prediction model individually, in order to carry out effective customer churn predictions in heterogeneous industries. Of course, in some studies, there are studies in which customers are subdivided using clustering techniques and applied a churn prediction model for individual customer groups. Although this process of predicting churn can produce better predictions than a single predict model for the entire customer population, there is still room for improvement in that clustering is a mechanical, exploratory grouping technique that calculates distances based on inputs and does not reflect the strategic intent of an entity such as loyalties. This study proposes a segment-based customer departure prediction process (CCP/2DL: Customer Churn Prediction based on Two-Dimensional Loyalty segmentation) based on two-dimensional customer loyalty, assuming that successful customer churn management can be better done through improvements in the overall process than through the performance of the model itself. CCP/2DL is a series of churn prediction processes that segment two-way, quantitative and qualitative loyalty-based customer, conduct secondary grouping of customer segments according to churn patterns, and then independently apply heterogeneous churn prediction models for each churn pattern group. Performance comparisons were performed with the most commonly applied the General churn prediction process and the Clustering-based churn prediction process to assess the relative excellence of the proposed churn prediction process. The General churn prediction process used in this study refers to the process of predicting a single group of customers simply intended to be predicted as a machine learning model, using the most commonly used churn predicting method. And the Clustering-based churn prediction process is a method of first using clustering techniques to segment customers and implement a churn prediction model for each individual group. In cooperation with a global NGO, the proposed CCP/2DL performance showed better performance than other methodologies for predicting churn. This churn prediction process is not only effective in predicting churn, but can also be a strategic basis for obtaining a variety of customer observations and carrying out other related performance marketing activities.

• 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.19 no.2
• /
• pp.39-54
• /
• 2013

Recent explosive increase of electronic commerce provides many advantageous purchase opportunities to customers. In this situation, customers who do not have enough knowledge about their purchases, may accept product recommendations. Product recommender systems automatically reflect user's preference and provide recommendation list to the users. Thus, product recommender system in online shopping store has been known as one of the most popular tools for one-to-one marketing. However, recommender systems which do not properly reflect user's preference cause user's disappointment and waste of time. In this study, we propose a novel recommender system which uses data mining and multi-model ensemble techniques to enhance the recommendation performance through reflecting the precise user's preference. The research data is collected from the real-world online shopping store, which deals products from famous art galleries and museums in Korea. The data initially contain 5759 transaction data, but finally remain 3167 transaction data after deletion of null data. In this study, we transform the categorical variables into dummy variables and exclude outlier data. The proposed model consists of two steps. The first step predicts customers who have high likelihood to purchase products in the online shopping store. In this step, we first use logistic regression, decision trees, and artificial neural networks to predict customers who have high likelihood to purchase products in each product group. We perform above data mining techniques using SAS E-Miner software. In this study, we partition datasets into two sets as modeling and validation sets for the logistic regression and decision trees. We also partition datasets into three sets as training, test, and validation sets for the artificial neural network model. The validation dataset is equal for the all experiments. Then we composite the results of each predictor using the multi-model ensemble techniques such as bagging and bumping. Bagging is the abbreviation of "Bootstrap Aggregation" and it composite outputs from several machine learning techniques for raising the performance and stability of prediction or classification. This technique is special form of the averaging method. Bumping is the abbreviation of "Bootstrap Umbrella of Model Parameter," and it only considers the model which has the lowest error value. The results show that bumping outperforms bagging and the other predictors except for "Poster" product group. For the "Poster" product group, artificial neural network model performs better than the other models. In the second step, we use the market basket analysis to extract association rules for co-purchased products. We can extract thirty one association rules according to values of Lift, Support, and Confidence measure. We set the minimum transaction frequency to support associations as 5%, maximum number of items in an association as 4, and minimum confidence for rule generation as 10%. This study also excludes the extracted association rules below 1 of lift value. We finally get fifteen association rules by excluding duplicate rules. Among the fifteen association rules, eleven rules contain association between products in "Office Supplies" product group, one rules include the association between "Office Supplies" and "Fashion" product groups, and other three rules contain association between "Office Supplies" and "Home Decoration" product groups. Finally, the proposed product recommender systems provides list of recommendations to the proper customers. We test the usability of the proposed system by using prototype and real-world transaction and profile data. For this end, we construct the prototype system by using the ASP, Java Script and Microsoft Access. In addition, we survey about user satisfaction for the recommended product list from the proposed system and the randomly selected product lists. The participants for the survey are 173 persons who use MSN Messenger, Daum Caf$\acute{e}$, and P2P services. We evaluate the user satisfaction using five-scale Likert measure. This study also performs "Paired Sample T-test" for the results of the survey. The results show that the proposed model outperforms the random selection model with 1% statistical significance level. It means that the users satisfied the recommended product list significantly. The results also show that the proposed system may be useful in real-world online shopping store.

• Korean Journal of Agricultural and Forest Meteorology
• /
• v.23 no.1
• /
• pp.15-33
• /
• 2021

In this study, the possibility of discriminating Fire blight (FB) infection tested using the hyperspectral imagery. The reflectance of healthy and infected leaves and branches was acquired with 5 nm of full width at high maximum (FWHM) and then it was standardized to 10 nm, 25 nm, 50 nm, and 80 nm of FWHM. The standardized samples were divided into training and test sets at ratios of 7:3, 5:5 and 3:7 to find the optimal bands of FWHM by the decision tree analysis. Classification accuracy was evaluated using overall accuracy (OA) and kappa coefficient (KC). The hyperspectral reflectance of infected leaves and branches was significantly lower than those of healthy green, red-edge (RE) and near infrared (NIR) regions. The bands selected for the first node were generally 750 and 800 nm; these were used to identify the infection of leaves and branches, respectively. The accuracy of the classifier was higher in the 7:3 ratio. Four bands with 50 nm of FWHM (450, 650, 750, and 950 nm) might be reasonable because the difference in the recalculated accuracy between 8 bands with 10 nm of FWHM (440, 580, 640, 660, 680, 710, 730, and 740 nm) and 4 bands was only 1.8% for OA and 4.1% for KC, respectively. Finally, adding two bands (550 nm and 800 nm with 25 nm of FWHM) in four bands with 50 nm of FWHM have been proposed to improve the usability of multispectral image sensors with performing various roles in agriculture as well as detecting FB with other combinations of spectral bands.

• Economic and Environmental Geology
• /
• v.55 no.5
• /
• pp.551-561
• /
• 2022

Geophysical exploration methods are very useful for generating high-resolution images of underground structures, and such methods can be applied to investigation of buried cultural properties and for determining their exact locations. In this study, image feature extraction and image segmentation methods were applied to automatically distinguish the structures of buried relics from the high-resolution ground-penetrating radar (GPR) images obtained at the center of Silla Kingdom, Gyeongju, South Korea. The major purpose for image feature extraction analyses is identifying the circular features from building remains and the linear features from ancient roads and fences. Feature extraction is implemented by applying the Canny edge detection and Hough transform algorithms. We applied the Hough transforms to the edge image resulted from the Canny algorithm in order to determine the locations the target features. However, the Hough transform requires different parameter settings for each survey sector. As for image segmentation, we applied the connected element labeling algorithm and object-based image analysis using Orfeo Toolbox (OTB) in QGIS. The connected components labeled image shows the signals associated with the target buried relics are effectively connected and labeled. However, we often find multiple labels are assigned to a single structure on the given GPR data. Object-based image analysis was conducted by using a Large-Scale Mean-Shift (LSMS) image segmentation. In this analysis, a vector layer containing pixel values for each segmented polygon was estimated first and then used to build a train-validation dataset by assigning the polygons to one class associated with the buried relics and another class for the background field. With the Random Forest Classifier, we find that the polygons on the LSMS image segmentation layer can be successfully classified into the polygons of the buried relics and those of the background. Thus, we propose that these automatic classification methods applied to the GPR images of buried cultural heritage in this study can be useful to obtain consistent analyses results for planning excavation processes.

• Asia pacific journal of information systems
• /
• v.21 no.1
• /
• pp.103-122
• /
• 2011

Recently, numerous documents have been made available electronically. Internet search engines and digital libraries commonly return query results containing hundreds or even thousands of documents. In this situation, it is virtually impossible for users to examine complete documents to determine whether they might be useful for them. For this reason, some on-line documents are accompanied by a list of keywords specified by the authors in an effort to guide the users by facilitating the filtering process. In this way, a set of keywords is often considered a condensed version of the whole document and therefore plays an important role for document retrieval, Web page retrieval, document clustering, summarization, text mining, and so on. Since many academic journals ask the authors to provide a list of five or six keywords on the first page of an article, keywords are most familiar in the context of journal articles. However, many other types of documents could not benefit from the use of keywords, including Web pages, email messages, news reports, magazine articles, and business papers. Although the potential benefit is large, the implementation itself is the obstacle; manually assigning keywords to all documents is a daunting task, or even impractical in that it is extremely tedious and time-consuming requiring a certain level of domain knowledge. Therefore, it is highly desirable to automate the keyword generation process. There are mainly two approaches to achieving this aim: keyword assignment approach and keyword extraction approach. Both approaches use machine learning methods and require, for training purposes, a set of documents with keywords already attached. In the former approach, there is a given set of vocabulary, and the aim is to match them to the texts. In other words, the keywords assignment approach seeks to select the words from a controlled vocabulary that best describes a document. Although this approach is domain dependent and is not easy to transfer and expand, it can generate implicit keywords that do not appear in a document. On the other hand, in the latter approach, the aim is to extract keywords with respect to their relevance in the text without prior vocabulary. In this approach, automatic keyword generation is treated as a classification task, and keywords are commonly extracted based on supervised learning techniques. Thus, keyword extraction algorithms classify candidate keywords in a document into positive or negative examples. Several systems such as Extractor and Kea were developed using keyword extraction approach. Most indicative words in a document are selected as keywords for that document and as a result, keywords extraction is limited to terms that appear in the document. Therefore, keywords extraction cannot generate implicit keywords that are not included in a document. According to the experiment results of Turney, about 64% to 90% of keywords assigned by the authors can be found in the full text of an article. Inversely, it also means that 10% to 36% of the keywords assigned by the authors do not appear in the article, which cannot be generated through keyword extraction algorithms. Our preliminary experiment result also shows that 37% of keywords assigned by the authors are not included in the full text. This is the reason why we have decided to adopt the keyword assignment approach. In this paper, we propose a new approach for automatic keyword assignment namely IVSM(Inverse Vector Space Model). The model is based on a vector space model. which is a conventional information retrieval model that represents documents and queries by vectors in a multidimensional space. IVSM generates an appropriate keyword set for a specific document by measuring the distance between the document and the keyword sets. The keyword assignment process of IVSM is as follows: (1) calculating the vector length of each keyword set based on each keyword weight; (2) preprocessing and parsing a target document that does not have keywords; (3) calculating the vector length of the target document based on the term frequency; (4) measuring the cosine similarity between each keyword set and the target document; and (5) generating keywords that have high similarity scores. Two keyword generation systems were implemented applying IVSM: IVSM system for Web-based community service and stand-alone IVSM system. Firstly, the IVSM system is implemented in a community service for sharing knowledge and opinions on current trends such as fashion, movies, social problems, and health information. The stand-alone IVSM system is dedicated to generating keywords for academic papers, and, indeed, it has been tested through a number of academic papers including those published by the Korean Association of Shipping and Logistics, the Korea Research Academy of Distribution Information, the Korea Logistics Society, the Korea Logistics Research Association, and the Korea Port Economic Association. We measured the performance of IVSM by the number of matches between the IVSM-generated keywords and the author-assigned keywords. According to our experiment, the precisions of IVSM applied to Web-based community service and academic journals were 0.75 and 0.71, respectively. The performance of both systems is much better than that of baseline systems that generate keywords based on simple probability. Also, IVSM shows comparable performance to Extractor that is a representative system of keyword extraction approach developed by Turney. As electronic documents increase, we expect that IVSM proposed in this paper can be applied to many electronic documents in Web-based community and digital library.

• Journal of Intelligence and Information Systems
• /
• v.21 no.2
• /
• pp.69-92
• /
• 2015

The explosion of social media data has led to apply text-mining techniques to analyze big social media data in a more rigorous manner. Even if social media text analysis algorithms were improved, previous approaches to social media text analysis have some limitations. In the field of sentiment analysis of social media written in Korean, there are two typical approaches. One is the linguistic approach using machine learning, which is the most common approach. Some studies have been conducted by adding grammatical factors to feature sets for training classification model. The other approach adopts the semantic analysis method to sentiment analysis, but this approach is mainly applied to English texts. To overcome these limitations, this study applies the Word2Vec algorithm which is an extension of the neural network algorithms to deal with more extensive semantic features that were underestimated in existing sentiment analysis. The result from adopting the Word2Vec algorithm is compared to the result from co-occurrence analysis to identify the difference between two approaches. The results show that the distribution related word extracted by Word2Vec algorithm in that the words represent some emotion about the keyword used are three times more than extracted by co-occurrence analysis. The reason of the difference between two results comes from Word2Vec's semantic features vectorization. Therefore, it is possible to say that Word2Vec algorithm is able to catch the hidden related words which have not been found in traditional analysis. In addition, Part Of Speech (POS) tagging for Korean is used to detect adjective as "emotional word" in Korean. In addition, the emotion words extracted from the text are converted into word vector by the Word2Vec algorithm to find related words. Among these related words, noun words are selected because each word of them would have causal relationship with "emotional word" in the sentence. The process of extracting these trigger factor of emotional word is named "Emotion Trigger" in this study. As a case study, the datasets used in the study are collected by searching using three keywords: professor, prosecutor, and doctor in that these keywords contain rich public emotion and opinion. Advanced data collecting was conducted to select secondary keywords for data gathering. The secondary keywords for each keyword used to gather the data to be used in actual analysis are followed: Professor (sexual assault, misappropriation of research money, recruitment irregularities, polifessor), Doctor (Shin hae-chul sky hospital, drinking and plastic surgery, rebate) Prosecutor (lewd behavior, sponsor). The size of the text data is about to 100,000(Professor: 25720, Doctor: 35110, Prosecutor: 43225) and the data are gathered from news, blog, and twitter to reflect various level of public emotion into text data analysis. As a visualization method, Gephi (http://gephi.github.io) was used and every program used in text processing and analysis are java coding. The contributions of this study are as follows: First, different approaches for sentiment analysis are integrated to overcome the limitations of existing approaches. Secondly, finding Emotion Trigger can detect the hidden connections to public emotion which existing method cannot detect. Finally, the approach used in this study could be generalized regardless of types of text data. The limitation of this study is that it is hard to say the word extracted by Emotion Trigger processing has significantly causal relationship with emotional word in a sentence. The future study will be conducted to clarify the causal relationship between emotional words and the words extracted by Emotion Trigger by comparing with the relationships manually tagged. Furthermore, the text data used in Emotion Trigger are twitter, so the data have a number of distinct features which we did not deal with in this study. These features will be considered in further study.

• Journal of Intelligence and Information Systems
• /
• v.26 no.1
• /
• pp.119-133
• /
• 2020

Fama asserted that in an efficient market, we can't make a trading rule that consistently outperforms the average stock market returns. This study aims to suggest a machine learning algorithm to improve the trading performance of an intraday short volatility strategy applying asymmetric volatility spillover effect, and analyze its trading performance improvement. Generally stock market volatility has a negative relation with stock market return and the Korean stock market volatility is influenced by the US stock market volatility. This volatility spillover effect is asymmetric. The asymmetric volatility spillover effect refers to the phenomenon that the US stock market volatility up and down differently influence the next day's volatility of the Korean stock market. We collected the S&P 500 index, VIX, KOSPI 200 index, and V-KOSPI 200 from 2008 to 2018. We found the negative relation between the S&P 500 and VIX, and the KOSPI 200 and V-KOSPI 200. We also documented the strong volatility spillover effect from the VIX to the V-KOSPI 200. Interestingly, the asymmetric volatility spillover was also found. Whereas the VIX up is fully reflected in the opening volatility of the V-KOSPI 200, the VIX down influences partially in the opening volatility and its influence lasts to the Korean market close. If the stock market is efficient, there is no reason why there exists the asymmetric volatility spillover effect. It is a counter example of the efficient market hypothesis. To utilize this type of anomalous volatility spillover pattern, we analyzed the intraday volatility selling strategy. This strategy sells short the Korean volatility market in the morning after the US stock market volatility closes down and takes no position in the volatility market after the VIX closes up. It produced profit every year between 2008 and 2018 and the percent profitable is 68%. The trading performance showed the higher average annual return of 129% relative to the benchmark average annual return of 33%. The maximum draw down, MDD, is -41%, which is lower than that of benchmark -101%. The Sharpe ratio 0.32 of SVS strategy is much greater than the Sharpe ratio 0.08 of the Benchmark strategy. The Sharpe ratio simultaneously considers return and risk and is calculated as return divided by risk. Therefore, high Sharpe ratio means high performance when comparing different strategies with different risk and return structure. Real world trading gives rise to the trading costs including brokerage cost and slippage cost. When the trading cost is considered, the performance difference between 76% and -10% average annual returns becomes clear. To improve the performance of the suggested volatility trading strategy, we used the well-known SVM algorithm. Input variables include the VIX close to close return at day t-1, the VIX open to close return at day t-1, the VK open return at day t, and output is the up and down classification of the VK open to close return at day t. The training period is from 2008 to 2014 and the testing period is from 2015 to 2018. The kernel functions are linear function, radial basis function, and polynomial function. We suggested the modified-short volatility strategy that sells the VK in the morning when the SVM output is Down and takes no position when the SVM output is Up. The trading performance was remarkably improved. The 5-year testing period trading results of the m-SVS strategy showed very high profit and low risk relative to the benchmark SVS strategy. The annual return of the m-SVS strategy is 123% and it is higher than that of SVS strategy. The risk factor, MDD, was also significantly improved from -41% to -29%.