• Journal of Intelligence and Information Systems
• /
• v.27 no.3
• /
• pp.57-73
• /
• 2021

Maintenance and prevention of failure through anomaly detection of ICT infrastructure is becoming important. System monitoring data is multidimensional time series data. When we deal with multidimensional time series data, we have difficulty in considering both characteristics of multidimensional data and characteristics of time series data. When dealing with multidimensional data, correlation between variables should be considered. Existing methods such as probability and linear base, distance base, etc. are degraded due to limitations called the curse of dimensions. In addition, time series data is preprocessed by applying sliding window technique and time series decomposition for self-correlation analysis. These techniques are the cause of increasing the dimension of data, so it is necessary to supplement them. The anomaly detection field is an old research field, and statistical methods and regression analysis were used in the early days. Currently, there are active studies to apply machine learning and artificial neural network technology to this field. Statistically based methods are difficult to apply when data is non-homogeneous, and do not detect local outliers well. The regression analysis method compares the predictive value and the actual value after learning the regression formula based on the parametric statistics and it detects abnormality. Anomaly detection using regression analysis has the disadvantage that the performance is lowered when the model is not solid and the noise or outliers of the data are included. There is a restriction that learning data with noise or outliers should be used. The autoencoder using artificial neural networks is learned to output as similar as possible to input data. It has many advantages compared to existing probability and linear model, cluster analysis, and map learning. It can be applied to data that does not satisfy probability distribution or linear assumption. In addition, it is possible to learn non-mapping without label data for teaching. However, there is a limitation of local outlier identification of multidimensional data in anomaly detection, and there is a problem that the dimension of data is greatly increased due to the characteristics of time series data. In this study, we propose a CMAE (Conditional Multimodal Autoencoder) that enhances the performance of anomaly detection by considering local outliers and time series characteristics. First, we applied Multimodal Autoencoder (MAE) to improve the limitations of local outlier identification of multidimensional data. Multimodals are commonly used to learn different types of inputs, such as voice and image. The different modal shares the bottleneck effect of Autoencoder and it learns correlation. In addition, CAE (Conditional Autoencoder) was used to learn the characteristics of time series data effectively without increasing the dimension of data. In general, conditional input mainly uses category variables, but in this study, time was used as a condition to learn periodicity. The CMAE model proposed in this paper was verified by comparing with the Unimodal Autoencoder (UAE) and Multi-modal Autoencoder (MAE). The restoration performance of Autoencoder for 41 variables was confirmed in the proposed model and the comparison model. The restoration performance is different by variables, and the restoration is normally well operated because the loss value is small for Memory, Disk, and Network modals in all three Autoencoder models. The process modal did not show a significant difference in all three models, and the CPU modal showed excellent performance in CMAE. ROC curve was prepared for the evaluation of anomaly detection performance in the proposed model and the comparison model, and AUC, accuracy, precision, recall, and F1-score were compared. In all indicators, the performance was shown in the order of CMAE, MAE, and AE. Especially, the reproduction rate was 0.9828 for CMAE, which can be confirmed to detect almost most of the abnormalities. The accuracy of the model was also improved and 87.12%, and the F1-score was 0.8883, which is considered to be suitable for anomaly detection. In practical aspect, the proposed model has an additional advantage in addition to performance improvement. The use of techniques such as time series decomposition and sliding windows has the disadvantage of managing unnecessary procedures; and their dimensional increase can cause a decrease in the computational speed in inference.The proposed model has characteristics that are easy to apply to practical tasks such as inference speed and model management.

• Journal of Intelligence and Information Systems
• /
• v.27 no.1
• /
• pp.83-102
• /
• 2021

The government recently announced various policies for developing big-data and artificial intelligence fields to provide a great opportunity to the public with respect to disclosure of high-quality data within public institutions. KSURE(Korea Trade Insurance Corporation) is a major public institution for financial policy in Korea, and thus the company is strongly committed to backing export companies with various systems. Nevertheless, there are still fewer cases of realized business model based on big-data analyses. In this situation, this paper aims to develop a new business model which can be applied to an ex-ante prediction for the likelihood of the insurance accident of credit guarantee. We utilize internal data from KSURE which supports export companies in Korea and apply machine learning models. Then, we conduct performance comparison among the predictive models including Logistic Regression, Random Forest, XGBoost, LightGBM, and DNN(Deep Neural Network). For decades, many researchers have tried to find better models which can help to predict bankruptcy since the ex-ante prediction is crucial for corporate managers, investors, creditors, and other stakeholders. The development of the prediction for financial distress or bankruptcy was originated from Smith(1930), Fitzpatrick(1932), or Merwin(1942). One of the most famous models is the Altman's Z-score model(Altman, 1968) which was based on the multiple discriminant analysis. This model is widely used in both research and practice by this time. The author suggests the score model that utilizes five key financial ratios to predict the probability of bankruptcy in the next two years. Ohlson(1980) introduces logit model to complement some limitations of previous models. Furthermore, Elmer and Borowski(1988) develop and examine a rule-based, automated system which conducts the financial analysis of savings and loans. Since the 1980s, researchers in Korea have started to examine analyses on the prediction of financial distress or bankruptcy. Kim(1987) analyzes financial ratios and develops the prediction model. Also, Han et al.(1995, 1996, 1997, 2003, 2005, 2006) construct the prediction model using various techniques including artificial neural network. Yang(1996) introduces multiple discriminant analysis and logit model. Besides, Kim and Kim(2001) utilize artificial neural network techniques for ex-ante prediction of insolvent enterprises. After that, many scholars have been trying to predict financial distress or bankruptcy more precisely based on diverse models such as Random Forest or SVM. One major distinction of our research from the previous research is that we focus on examining the predicted probability of default for each sample case, not only on investigating the classification accuracy of each model for the entire sample. Most predictive models in this paper show that the level of the accuracy of classification is about 70% based on the entire sample. To be specific, LightGBM model shows the highest accuracy of 71.1% and Logit model indicates the lowest accuracy of 69%. However, we confirm that there are open to multiple interpretations. In the context of the business, we have to put more emphasis on efforts to minimize type 2 error which causes more harmful operating losses for the guaranty company. Thus, we also compare the classification accuracy by splitting predicted probability of the default into ten equal intervals. When we examine the classification accuracy for each interval, Logit model has the highest accuracy of 100% for 0~10% of the predicted probability of the default, however, Logit model has a relatively lower accuracy of 61.5% for 90~100% of the predicted probability of the default. On the other hand, Random Forest, XGBoost, LightGBM, and DNN indicate more desirable results since they indicate a higher level of accuracy for both 0~10% and 90~100% of the predicted probability of the default but have a lower level of accuracy around 50% of the predicted probability of the default. When it comes to the distribution of samples for each predicted probability of the default, both LightGBM and XGBoost models have a relatively large number of samples for both 0~10% and 90~100% of the predicted probability of the default. Although Random Forest model has an advantage with regard to the perspective of classification accuracy with small number of cases, LightGBM or XGBoost could become a more desirable model since they classify large number of cases into the two extreme intervals of the predicted probability of the default, even allowing for their relatively low classification accuracy. Considering the importance of type 2 error and total prediction accuracy, XGBoost and DNN show superior performance. Next, Random Forest and LightGBM show good results, but logistic regression shows the worst performance. However, each predictive model has a comparative advantage in terms of various evaluation standards. For instance, Random Forest model shows almost 100% accuracy for samples which are expected to have a high level of the probability of default. Collectively, we can construct more comprehensive ensemble models which contain multiple classification machine learning models and conduct majority voting for maximizing its overall performance.

• Cartoon and Animation Studies
• /
• s.48
• /
• pp.85-116
• /
• 2017

The purpose of this study is to analyze conditions that robots depicted in cinematographic works like animations or movies sympathize with and form an attachment with the nuclear person and organize characterizing strategies for emotional sympathetic robots. Along with the development of technology, the areas of artificial intelligence and robots are no longer considered to belong to science fiction but as realistic issues. Therefore, this author assumes that the expressive characteristics of emotional sympathetic robots created by cinematographic works should be used as meaningful factors in expressively embodying human-friendly service robots to be distributed widely afterwards, that is, in establishing the features of characters. To lay the grounds for it, this research has begun. As the subjects of analysis, this researcher has chosen robot characters whose emotional intimacy with the main person is clearly observed among those found in movies and animations produced after the 1920 when robot's contemporary concept was declared. Also, to understand robots' appearance and behavioral tendency, this study (1) has classified robots' external impressions into five types (human-like, cartoon, tool-like, artificial bring, pet or creature) and (2) has classified behavioral tendencies considered to be the outer embodiment of personality by using DiSC, the tool to diagnose behavioral patterns. Meanwhile, it has been observed that robots equipped with high emotional intimacy are all strongly independent about their duties and indicate great emotional acceptance. Therefore, 'influence' and 'Steadiness' types show great emotional acceptance, the influencing type tends to be highly independent, and the 'Conscientiousness' type tends to indicate less emotional acceptance and independency in general. Yet, according to the analysis on external impressions, appearance factors hardly have any significant relationship with emotional sympathy. It implies that regarding the conditions of robots equipped with great emotional sympathy, emotional sympathy grounded on communication exerts more crucial effects than first impression similarly to the process of forming interpersonal relationship in reality. Lastly, to study the characters of robots, it is absolutely needed to have consilient competence embracing different areas widely. This author also has felt that only with design factors or personality factors, it is hard to estimate robot characters and also analyze a vast amount of information demanded in sympathy with humans entirely. However, this researcher will end this thesis as the foundation for it expecting that the general artistic value of animations can be used preciously afterwards in developing robots that have to be studied interdisciplinarily.

• Journal of Korea Water Resources Association
• /
• v.54 no.10
• /
• pp.795-805
• /
• 2021

In this study, after developing an LSTM-based deep learning model for estimating daily runoff in the Soyang River Dam basin, the accuracy of the model for various combinations of model structure and input data was investigated. A model was built based on the database consisting of average daily precipitation, average daily temperature, average daily wind speed (input up to here), and daily average flow rate (output) during the first 12 years (1997.1.1-2008.12.31). The Nash-Sutcliffe Model Efficiency Coefficient (NSE) and RMSE were examined for validation using the flow discharge data of the later 12 years (2009.1.1-2020.12.31). The combination that showed the highest accuracy was the case in which all possible input data (12 years of daily precipitation, weather temperature, wind speed) were used on the LSTM model structure with 64 hidden units. The NSE and RMSE of the verification period were 0.862 and 76.8 m³/s, respectively. When the number of hidden units of LSTM exceeds 500, the performance degradation of the model due to overfitting begins to appear, and when the number of hidden units exceeds 1000, the overfitting problem becomes prominent. A model with very high performance (NSE=0.8~0.84) could be obtained when only 12 years of daily precipitation was used for model training. A model with reasonably high performance (NSE=0.63-0.85) when only one year of input data was used for model training. In particular, an accurate model (NSE=0.85) could be obtained if the one year of training data contains a wide magnitude of flow events such as extreme flow and droughts as well as normal events. If the training data includes both the normal and extreme flow rates, input data that is longer than 5 years did not significantly improve the model performance.

• Tunnel and Underground Space
• /
• v.31 no.5
• /
• pp.374-384
• /
• 2021

In this study, we propose a new approach for automatic fracture detection in CT scan images of rock specimens. This approach is built on top of two-stage object detection deep learning algorithm called Faster R-CNN with a major modification of using rotated bounding box. The use of rotated bounding box plays a key role in the future work to overcome several inherent difficulties of fracture segmentation relating to the heterogeneity of uninterested background (i.e., minerals) and the variation in size and shape of fracture. Comparing to the commonly used bounding box (i.e., axis-align bounding box), rotated bounding box shows a greater adaptability to fit with the elongated shape of fracture, such that minimizing the ratio of background within the bounding box. Besides, an additional benefit of rotated bounding box is that it can provide relative information on the orientation and length of fracture without the further segmentation and measurement step. To validate the applicability of the proposed approach, we train and test our approach with a number of CT image sets of fractured granite specimens with highly heterogeneous background and other rocks such as sandstone and shale. The result demonstrates that our approach can lead to the encouraging results on fracture detection with the mean average precision (mAP) up to 0.89 and also outperform the conventional approach in terms of background-to-object ratio within the bounding box.

• Journal of Intelligence and Information Systems
• /
• v.22 no.1
• /
• pp.187-204
• /
• 2016

Document classification based on emotional polarity has become a welcomed emerging task owing to the great explosion of data on the Web. In the big data age, there are too many information sources to refer to when making decisions. For example, when considering travel to a city, a person may search reviews from a search engine such as Google or social networking services (SNSs) such as blogs, Twitter, and Facebook. The emotional polarity of positive and negative reviews helps a user decide on whether or not to make a trip. Sentiment analysis of customer reviews has become an important research topic as datamining technology is widely accepted for text mining of the Web. Sentiment analysis has been used to classify documents through machine learning techniques, such as the decision tree, neural networks, and support vector machines (SVMs). is used to determine the attitude, position, and sensibility of people who write articles about various topics that are published on the Web. Regardless of the polarity of customer reviews, emotional reviews are very helpful materials for analyzing the opinions of customers through their reviews. Sentiment analysis helps with understanding what customers really want instantly through the help of automated text mining techniques. Sensitivity analysis utilizes text mining techniques on text on the Web to extract subjective information in the text for text analysis. Sensitivity analysis is utilized to determine the attitudes or positions of the person who wrote the article and presented their opinion about a particular topic. In this study, we developed a model that selects a hot topic from user posts at China's online stock forum by using the k-means algorithm and self-organizing map (SOM). In addition, we developed a detecting model to predict a hot topic by using machine learning techniques such as logit, the decision tree, and SVM. We employed sensitivity analysis to develop our model for the selection and detection of hot topics from China's online stock forum. The sensitivity analysis calculates a sentimental value from a document based on contrast and classification according to the polarity sentimental dictionary (positive or negative). The online stock forum was an attractive site because of its information about stock investment. Users post numerous texts about stock movement by analyzing the market according to government policy announcements, market reports, reports from research institutes on the economy, and even rumors. We divided the online forum's topics into 21 categories to utilize sentiment analysis. One hundred forty-four topics were selected among 21 categories at online forums about stock. The posts were crawled to build a positive and negative text database. We ultimately obtained 21,141 posts on 88 topics by preprocessing the text from March 2013 to February 2015. The interest index was defined to select the hot topics, and the k-means algorithm and SOM presented equivalent results with this data. We developed a decision tree model to detect hot topics with three algorithms: CHAID, CART, and C4.5. The results of CHAID were subpar compared to the others. We also employed SVM to detect the hot topics from negative data. The SVM models were trained with the radial basis function (RBF) kernel function by a grid search to detect the hot topics. The detection of hot topics by using sentiment analysis provides the latest trends and hot topics in the stock forum for investors so that they no longer need to search the vast amounts of information on the Web. Our proposed model is also helpful to rapidly determine customers' signals or attitudes towards government policy and firms' products and services.

• Journal of Intelligence and Information Systems
• /
• v.22 no.4
• /
• pp.177-192
• /
• 2016

Machine learning is a field of artificial intelligence. It refers to an area of computer science related to providing machines the ability to perform their own data analysis, decision making and forecasting. For example, one of the representative machine learning models is artificial neural network, which is a statistical learning algorithm inspired by the neural network structure of biology. In addition, there are other machine learning models such as decision tree model, naive bayes model and SVM(support vector machine) model. Among the machine learning models, we use SVM model in this study because it is mainly used for classification and regression analysis that fits well to our study. The core principle of SVM is to find a reasonable hyperplane that distinguishes different group in the data space. Given information about the data in any two groups, the SVM model judges to which group the new data belongs based on the hyperplane obtained from the given data set. Thus, the more the amount of meaningful data, the better the machine learning ability. In recent years, many financial experts have focused on machine learning, seeing the possibility of combining with machine learning and the financial field where vast amounts of financial data exist. Machine learning techniques have been proved to be powerful in describing the non-stationary and chaotic stock price dynamics. A lot of researches have been successfully conducted on forecasting of stock prices using machine learning algorithms. Recently, financial companies have begun to provide Robo-Advisor service, a compound word of Robot and Advisor, which can perform various financial tasks through advanced algorithms using rapidly changing huge amount of data. Robo-Adviser's main task is to advise the investors about the investor's personal investment propensity and to provide the service to manage the portfolio automatically. In this study, we propose a method of forecasting the Korean volatility index, VKOSPI, using the SVM model, which is one of the machine learning methods, and applying it to real option trading to increase the trading performance. VKOSPI is a measure of the future volatility of the KOSPI 200 index based on KOSPI 200 index option prices. VKOSPI is similar to the VIX index, which is based on S&P 500 option price in the United States. The Korea Exchange(KRX) calculates and announce the real-time VKOSPI index. VKOSPI is the same as the usual volatility and affects the option prices. The direction of VKOSPI and option prices show positive relation regardless of the option type (call and put options with various striking prices). If the volatility increases, all of the call and put option premium increases because the probability of the option's exercise possibility increases. The investor can know the rising value of the option price with respect to the volatility rising value in real time through Vega, a Black-Scholes's measurement index of an option's sensitivity to changes in the volatility. Therefore, accurate forecasting of VKOSPI movements is one of the important factors that can generate profit in option trading. In this study, we verified through real option data that the accurate forecast of VKOSPI is able to make a big profit in real option trading. To the best of our knowledge, there have been no studies on the idea of predicting the direction of VKOSPI based on machine learning and introducing the idea of applying it to actual option trading. In this study predicted daily VKOSPI changes through SVM model and then made intraday option strangle position, which gives profit as option prices reduce, only when VKOSPI is expected to decline during daytime. We analyzed the results and tested whether it is applicable to real option trading based on SVM's prediction. The results showed the prediction accuracy of VKOSPI was 57.83% on average, and the number of position entry times was 43.2 times, which is less than half of the benchmark (100 times). A small number of trading is an indicator of trading efficiency. In addition, the experiment proved that the trading performance was significantly higher than the benchmark.

• Journal of Intelligence and Information Systems
• /
• v.21 no.1
• /
• pp.29-45
• /
• 2015

Response modeling is a well-known research issue for those who have tried to get more superior performance in the capability of predicting the customers' response for the marketing promotion. The response model for customers would reduce the marketing cost by identifying prospective customers from very large customer database and predicting the purchasing intention of the selected customers while the promotion which is derived from an undifferentiated marketing strategy results in unnecessary cost. In addition, the big data environment has accelerated developing the response model with data mining techniques such as CBR, neural networks and support vector machines. And CBR is one of the most major tools in business because it is known as simple and robust to apply to the response model. However, CBR is an attractive data mining technique for data mining applications in business even though it hasn't shown high performance compared to other machine learning techniques. Thus many studies have tried to improve CBR and utilized in business data mining with the enhanced algorithms or the support of other techniques such as genetic algorithm, decision tree and AHP (Analytic Process Hierarchy). Ahn and Kim(2008) utilized logit, neural networks, CBR to predict that which customers would purchase the items promoted by marketing department and tried to optimized the number of k for k-nearest neighbor with genetic algorithm for the purpose of improving the performance of the integrated model. Hong and Park(2009) noted that the integrated approach with CBR for logit, neural networks, and Support Vector Machine (SVM) showed more improved prediction ability for response of customers to marketing promotion than each data mining models such as logit, neural networks, and SVM. This paper presented an approach to predict customers' response of marketing promotion with Case Based Reasoning. The proposed model was developed by applying different weights to each feature. We deployed logit model with a database including the promotion and the purchasing data of bath soap. After that, the coefficients were used to give different weights of CBR. We analyzed the performance of proposed weighted CBR based model compared to neural networks and pure CBR based model empirically and found that the proposed weighted CBR based model showed more superior performance than pure CBR model. Imbalanced data is a common problem to build data mining model to classify a class with real data such as bankruptcy prediction, intrusion detection, fraud detection, churn management, and response modeling. Imbalanced data means that the number of instance in one class is remarkably small or large compared to the number of instance in other classes. The classification model such as response modeling has a lot of trouble to recognize the pattern from data through learning because the model tends to ignore a small number of classes while classifying a large number of classes correctly. To resolve the problem caused from imbalanced data distribution, sampling method is one of the most representative approach. The sampling method could be categorized to under sampling and over sampling. However, CBR is not sensitive to data distribution because it doesn't learn from data unlike machine learning algorithm. In this study, we investigated the robustness of our proposed model while changing the ratio of response customers and nonresponse customers to the promotion program because the response customers for the suggested promotion is always a small part of nonresponse customers in the real world. We simulated the proposed model 100 times to validate the robustness with different ratio of response customers to response customers under the imbalanced data distribution. Finally, we found that our proposed CBR based model showed superior performance than compared models under the imbalanced data sets. Our study is expected to improve the performance of response model for the promotion program with CBR under imbalanced data distribution in the real world.

• Journal of Intelligence and Information Systems
• /
• v.16 no.2
• /
• pp.19-32
• /
• 2010

Volatility plays a central role in both academic and practical applications, especially in pricing financial derivative products and trading volatility strategies. This study presents a novel mechanism based on generalized autoregressive conditional heteroskedasticity (GARCH) models that is able to enhance the performance of intelligent volatility trading systems by predicting Korean stock market volatility more accurately. In particular, we embedded the concept of the volatility asymmetry documented widely in the literature into our model. The newly developed Korean stock market volatility index of KOSPI 200, VKOSPI, is used as a volatility proxy. It is the price of a linear portfolio of the KOSPI 200 index options and measures the effect of the expectations of dealers and option traders on stock market volatility for 30 calendar days. The KOSPI 200 index options market started in 1997 and has become the most actively traded market in the world. Its trading volume is more than 10 million contracts a day and records the highest of all the stock index option markets. Therefore, analyzing the VKOSPI has great importance in understanding volatility inherent in option prices and can afford some trading ideas for futures and option dealers. Use of the VKOSPI as volatility proxy avoids statistical estimation problems associated with other measures of volatility since the VKOSPI is model-free expected volatility of market participants calculated directly from the transacted option prices. This study estimates the symmetric and asymmetric GARCH models for the KOSPI 200 index from January 2003 to December 2006 by the maximum likelihood procedure. Asymmetric GARCH models include GJR-GARCH model of Glosten, Jagannathan and Runke, exponential GARCH model of Nelson and power autoregressive conditional heteroskedasticity (ARCH) of Ding, Granger and Engle. Symmetric GARCH model indicates basic GARCH (1, 1). Tomorrow's forecasted value and change direction of stock market volatility are obtained by recursive GARCH specifications from January 2007 to December 2009 and are compared with the VKOSPI. Empirical results indicate that negative unanticipated returns increase volatility more than positive return shocks of equal magnitude decrease volatility, indicating the existence of volatility asymmetry in the Korean stock market. The point value and change direction of tomorrow VKOSPI are estimated and forecasted by GARCH models. Volatility trading system is developed using the forecasted change direction of the VKOSPI, that is, if tomorrow VKOSPI is expected to rise, a long straddle or strangle position is established. A short straddle or strangle position is taken if VKOSPI is expected to fall tomorrow. Total profit is calculated as the cumulative sum of the VKOSPI percentage change. If forecasted direction is correct, the absolute value of the VKOSPI percentage changes is added to trading profit. It is subtracted from the trading profit if forecasted direction is not correct. For the in-sample period, the power ARCH model best fits in a statistical metric, Mean Squared Prediction Error (MSPE), and the exponential GARCH model shows the highest Mean Correct Prediction (MCP). The power ARCH model best fits also for the out-of-sample period and provides the highest probability for the VKOSPI change direction tomorrow. Generally, the power ARCH model shows the best fit for the VKOSPI. All the GARCH models provide trading profits for volatility trading system and the exponential GARCH model shows the best performance, annual profit of 197.56%, during the in-sample period. The GARCH models present trading profits during the out-of-sample period except for the exponential GARCH model. During the out-of-sample period, the power ARCH model shows the largest annual trading profit of 38%. The volatility clustering and asymmetry found in this research are the reflection of volatility non-linearity. This further suggests that combining the asymmetric GARCH models and artificial neural networks can significantly enhance the performance of the suggested volatility trading system, since artificial neural networks have been shown to effectively model nonlinear relationships.

• Journal of Intelligence and Information Systems
• /
• v.25 no.2
• /
• pp.141-166
• /
• 2019

Recently, channels like social media and SNS create enormous amount of data. In all kinds of data, portions of unstructured data which represented as text data has increased geometrically. But there are some difficulties to check all text data, so it is important to access those data rapidly and grasp key points of text. Due to needs of efficient understanding, many studies about text summarization for handling and using tremendous amounts of text data have been proposed. Especially, a lot of summarization methods using machine learning and artificial intelligence algorithms have been proposed lately to generate summary objectively and effectively which called "automatic summarization". However almost text summarization methods proposed up to date construct summary focused on frequency of contents in original documents. Those summaries have a limitation for contain small-weight subjects that mentioned less in original text. If summaries include contents with only major subject, bias occurs and it causes loss of information so that it is hard to ascertain every subject documents have. To avoid those bias, it is possible to summarize in point of balance between topics document have so all subject in document can be ascertained, but still unbalance of distribution between those subjects remains. To retain balance of subjects in summary, it is necessary to consider proportion of every subject documents originally have and also allocate the portion of subjects equally so that even sentences of minor subjects can be included in summary sufficiently. In this study, we propose "subject-balanced" text summarization method that procure balance between all subjects and minimize omission of low-frequency subjects. For subject-balanced summary, we use two concept of summary evaluation metrics "completeness" and "succinctness". Completeness is the feature that summary should include contents of original documents fully and succinctness means summary has minimum duplication with contents in itself. Proposed method has 3-phases for summarization. First phase is constructing subject term dictionaries. Topic modeling is used for calculating topic-term weight which indicates degrees that each terms are related to each topic. From derived weight, it is possible to figure out highly related terms for every topic and subjects of documents can be found from various topic composed similar meaning terms. And then, few terms are selected which represent subject well. In this method, it is called "seed terms". However, those terms are too small to explain each subject enough, so sufficient similar terms with seed terms are needed for well-constructed subject dictionary. Word2Vec is used for word expansion, finds similar terms with seed terms. Word vectors are created after Word2Vec modeling, and from those vectors, similarity between all terms can be derived by using cosine-similarity. Higher cosine similarity between two terms calculated, higher relationship between two terms defined. So terms that have high similarity values with seed terms for each subjects are selected and filtering those expanded terms subject dictionary is finally constructed. Next phase is allocating subjects to every sentences which original documents have. To grasp contents of all sentences first, frequency analysis is conducted with specific terms that subject dictionaries compose. TF-IDF weight of each subjects are calculated after frequency analysis, and it is possible to figure out how much sentences are explaining about each subjects. However, TF-IDF weight has limitation that the weight can be increased infinitely, so by normalizing TF-IDF weights for every subject sentences have, all values are changed to 0 to 1 values. Then allocating subject for every sentences with maximum TF-IDF weight between all subjects, sentence group are constructed for each subjects finally. Last phase is summary generation parts. Sen2Vec is used to figure out similarity between subject-sentences, and similarity matrix can be formed. By repetitive sentences selecting, it is possible to generate summary that include contents of original documents fully and minimize duplication in summary itself. For evaluation of proposed method, 50,000 reviews of TripAdvisor are used for constructing subject dictionaries and 23,087 reviews are used for generating summary. Also comparison between proposed method summary and frequency-based summary is performed and as a result, it is verified that summary from proposed method can retain balance of all subject more which documents originally have.