• Journal of Internet Computing and Services
• /
• v.25 no.3
• /
• pp.43-54
• /
• 2024

This study employs text mining techniques to compare user responses across various Over-The-Top (OTT) service platforms. The primary objective of the research is to understand user satisfaction with OTT service platforms and contribute to the formulation of more effective review strategies. The key questions addressed in this study involve identifying prominent topics and keywords in user reviews of different OTT services and comprehending platform-specific user reactions. TF-IDF is utilized to extract significant words from positive and negative reviews, while BERTopic, an advanced topic modeling technique, is employed for a more nuanced and comprehensive analysis of intricate user reviews. The results from TF-IDF analysis reveal that positive app reviews exhibit a high frequency of content-related words, whereas negative reviews display a high frequency of words associated with potential issues during app usage. Through the utilization of BERTopic, we were able to extract keywords related to content diversity, app performance components, payment, and compatibility, by associating them with content attributes. This enabled us to verify that the distinguishing attributes of the platforms vary among themselves. The findings of this study offer significant insights into user behavior and preferences, which OTT service providers can leverage to improve user experience and satisfaction. We also anticipate that researchers exploring deep learning models will find our study results valuable for conducting analyses on user review text data.

• The Korean Journal of Applied Statistics
• /
• v.37 no.4
• /
• pp.525-539
• /
• 2024

Massive multiplayer online role playing game (MMORPG) is a common type of game these days. Predicting user churn in MMORPG is a crucial task. The retention rate of users is deeply associated with the lifespan and revenue of the service. If the churn of a specific user can be predicted in advance, targeted promotions can be used to encourage their stay. Therefore, not only the accuracy of churn prediction but also the speed at which signs of churn can be detected is important. In this paper, we propose methods to identify early signs of churn by utilizing the daily predicted user retention probabilities. We train various deep learning and machine learning models using log data and estimate user retention probabilities. By analyzing the change patterns in these probabilities, we provide empirical rules for early identification of users at high risk of churn. Performance evaluations confirm that our methodology is more effective at detecting high risk users than existing methods based on login days. Finally, we suggest novel methods for customized marketing strategies. For this purpose, we provide guidelines of the percentage of accessed users who are at risk of churn.

• Journal of Advanced Navigation Technology
• /
• v.28 no.5
• /
• pp.760-766
• /
• 2024

Machine learning and deep learning are being researched in various fields and applied in real life. Designing reliable models is crucial, and understanding the results of these models is necessary. This paper analyzes the impact of variables on prediction values using SHAP. Prediction models for PM_2.5 were designed using DNN and LSTM algorithms. The training and test data were composed by selecting weather data and air pollutant data through correlation analysis. The RMSE and accuracy for AQI categories were checked for both prediction models, with the LSTM algorithm showing slightly better performance. The contribution of variables to the prediction values of both models was confirmed using SHAP. It was found that air pollutant data had a high contribution in predicting PM_2.5, and temperature among weather data had a high contribution in the prediction process of both models. Both models showed that high values of temperature, wind speed, and sea level pressure decreased prediction values, while low values increased them. For NO₂ , PM₁₀, and SO₂, the LSTM model showed a bidirectional impact on prediction values, unlike the DNN model.

• KIPS Transactions on Computer and Communication Systems
• /
• v.10 no.10
• /
• pp.285-290
• /
• 2021

In the medical field, disease diagnosis and prediction research using artificial intelligence technology is being actively conducted. It is being released as a variety of products for disease diagnosis and prediction, which are most widely used in the application of artificial intelligence technology based on medical images. Artificial intelligence is being applied to diagnose diseases, to classify diseases into benign and malignant, and to separate disease regions for use in identification or reading according to the risk of disease. Recently, in connection with cloud technology, its utility as a service product is increasing. Among the diseases dealt with in this paper, liver disease is a disease with very high risk because it is difficult to diagnose early due to the lack of pain. Artificial intelligence technology was introduced based on medical images as a non-invasive diagnostic method for diagnosing these diseases. We describe the development of a web service to help the most meaningful clinical reading of liver cirrhosis patients. Then, it shows the web service process and shows the operation screen of each process and the final result screen. It is expected that the proposed service will be able to diagnose liver cirrhosis at an early stage and help patients recover through rapid treatment.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.30 no.6
• /
• pp.1131-1139
• /
• 2020

The static analysis of the source code is to find the remaining security weaknesses for a wide range of source codes. The static analysis tool is used to check the result, and the static analysis expert performs spying and false detection analysis on the result. In this process, the amount of analysis is large and the rate of false positives is high, so a lot of time and effort is required, and a method of efficient analysis is required. In addition, it is rare for experts to analyze only the source code of the line where the defect occurred when performing positive/false detection analysis. Depending on the type of defect, the surrounding source code is analyzed together and the final analysis result is delivered. In order to solve the difficulty of experts discriminating positive and false positives using these static analysis tools, this paper proposes a method of determining whether or not the security weakness found by the static analysis tools is a spy detection through artificial intelligence rather than an expert. In addition, the optimal size was confirmed through an experiment to see how the size of the training data (source code around the defects) used for such machine learning affects the performance. This result is expected to help the static analysis expert's job of classifying positive and false positives after static analysis.

• Journal of Korean Society of Disaster and Security
• /
• v.14 no.4
• /
• pp.17-27
• /
• 2021

Streamflow prediction is a very vital disaster mitigation approach for effective flood management and water resources planning. Lately, torrential rainfall caused by climate change has been reported to have increased globally, thereby causing enormous infrastructural loss, properties and lives. This study evaluates the contribution of rainfall to streamflow prediction in normal and peak rainfall scenarios, typical of the recent flood at Piney Resort in Vernon, Hickman County, Tennessee, United States. Daily streamflow, water level, and rainfall data for 20 years (2000-2019) from two USGS gage stations (03602500 upstream and 03599500 downstream) of the Piney River watershed were obtained, preprocesssed and fitted with Long short term memory (LSTM) model. Tensorflow and Keras machine learning frameworks were used with Python to predict streamflow values with a sequence size of 14 days, to determine whether the model could have predicted the flooding event in August 21, 2021. Model skill analysis showed that LSTM model with full data (water level, streamflow and rainfall) performed better than the Naive Model except some rainfall models, indicating that only rainfall is insufficient for streamflow prediction. The final LSTM model recorded optimal NSE and RMSE values of 0.68 and 13.84 m³/s and predicted peak flow with the lowest prediction error of 11.6%, indicating that the final model could have predicted the flood on August 24, 2021 given a peak rainfall scenario. Adequate knowledge of rainfall patterns will guide hydrologists and disaster prevention managers in designing efficient early warning systems and policies aimed at mitigating flood risks.

• Korean Journal of Cognitive Science
• /
• v.26 no.4
• /
• pp.377-392
• /
• 2015

Korean semantic role labeling (SRL) is usually performed by a machine learning and requires a lot of corpus. However, the Korean PropBank used in Korean SRL system is less than PropBank. It leads to a low performance. Therefore, we expand the annotated corpus and verb frames for Korean SRL system to expand the Korean PropBank corpus. Most of the SRL system have a domain-dependent performance so, the performance may decrease if domain was changed. In this paper, we use the domain adaptation technique to reduce decreasing performance with the existing corpus and the small size of new domain corpus. We apply the domain adaptation technique to Structural SVM and Deep Neural Network. The experimental result show the effectiveness of the domain adaptation technique.

• Journal of the Korea Convergence Society
• /
• v.10 no.9
• /
• pp.23-29
• /
• 2019

In this paper, the mood of images was classified into eight categories through a deep convolutional neural network and video was automatically generated using proper background music. Based on the collected image data, the classification model is learned using a multilayer perceptron (MLP). Using the MLP, a video is generated by using multi-class classification to predict image mood to be used for video generation, and by matching pre-classified music. As a result of 10-fold cross-validation and result of experiments on actual images, each 72.4% of accuracy and 64% of confusion matrix accuracy was achieved. In the case of misclassification, by classifying video into a similar mood, it was confirmed that the music from the video had no great mismatch with images.

• Database Research
• /
• v.34 no.3
• /
• pp.124-136
• /
• 2018

There have been active research activities to use neural networks to analyze OCT images and make medical decisions. One requirement for these approaches to be promising solutions is that the trained network must be generalized to new devices without a substantial loss of performance. In this paper, we use a deep convolutional neural network to distinguish AMD from normal patients. The network was trained using a data set generated from an OCT device. We observed a significant performance degradation when it was applied to a new data set obtained from a different OCT device. To overcome this performance degradation, we propose an image normalization method which performs segmentation of OCT images to identify the retina area and aligns images so that the retina region lies horizontally in the image. We experimentally evaluated the performance of the proposed method. The experiment confirmed a significant performance improvement of our approach.

• Journal of the Architectural Institute of Korea Planning & Design
• /
• v.34 no.12
• /
• pp.85-94
• /
• 2018

The purpose of this paper was to propose a model that recognizes potential users' emotional response toward design by classifying Electroencephalography(EEG). Studies in neuroscience and psychology have made an effort to recognize subjects' emotional response by analyzing EEG data. And this approach has been adopted in design since it is critical to monitor users' subjective response in the preface of design. Moreover, the building design process cannot be reversed after construction, recognizing clients' affection toward design alternatives plays important role. An experiment was conducted to record subjects' EEG data while they view their most/least liked images of small-house designs selected by them among the eight given images. After the recording, a subjective questionnaire, PANAS, was distributed to the subjects in order to describe their own affection score in quantitative way. Google TensorFlow was used to build and train the model. Dataset for model training and testing consist of feature columns for recorded EEG data and labels for the questionnaire results. After training and testing, the measured accuracy of the model was 0.975 which was higher than the other machine learning based classification methods. The proposed model may suggest one quantitative way of evaluating design alternatives. In addition, this method may support designer while designing the facilities for people like disabled or children who are not able to express their own feelings toward alternatives.