User behavior modeling in online games using machine learning techniques : focused on malicious user detection and churn prediction
- 주제(키워드) machine learning , malicious user detection , churn prediction , behavior modeling
- 발행기관 고려대학교 정보보호대학원
- 지도교수 김휘강
- 발행년도 2019
- 학위수여년월 2019. 8
- 유형 Text
- 학위구분 박사
- 학과 정보보호대학원 정보보호학과
- 원문페이지 126 p
- 실제URI http://www.dcollection.net/handler/korea/000000084744
- UCI I804:11009-000000084744
- DOI 10.23186/korea.000000084744.11009.0000932
- 본문언어 영어
- 제출원본 000045999333
초록/요약
Recently, machine learning has been widely used in various fields. Online game company also is trying to apply machine learning techniques to solve the problems that arise during the development and operation of online games. While these attempts are mainly focused on suggesting machine learning algorithms for efficient learning, there are other important issues that need to be addressed to apply machine learning in practice: Constructing training data, feature engineering, designing appropriate evaluation schemes, and maintaining the predictive model. This thesis proposes methodologies for solving the four problems described above when building a service based on a machine learning model in an online game. The methodologies was applied to live games which are serviced in NCSOFT. However, it still has some limitations. Recently, advances in machine learning have led to alternatives to overcome these limitations. Online games are a great platform to observe the behavior of various users in the virtual-world. In the virtual-world, a wide range of data is generated, ranging from activities performed by users, social relationships, economic systems, and even text data via chat between users. In other words, almost all forms of data that can be handled in the field of machine learning are generated with massive amounts. However, there is a few researches because of the lack of interchange between industry and academia. In this thesis, various issues and solutions are discussed to apply machine learning to live services in online games. I hope this study contributes a little to overcome the obstacles mentioned above.
more목차
Chapter 1. Introduction . １
Chapter 2. Game bot detection using self-similarity . ４
2.1. Introduction ４
2.2. Related works ６
2.3. Back ground . ８
A. Terminology . ８
B. Ecology of game world in MMORPGs . １０
C. Game bots and RMT １１
2.4. Game bot characteristics . １２
A. Exploratory data analysis １２
B. Feature review １９
2.5. Methodology ２１
A. Feature extraction and selection . ２１
B. Modeling and evaluation . ２３
C. Monitoring and retraining detection model . ２４
2.6. Experiments . ２６
A. Data preprocessing . ２６
B. Ground truth construction . ２７
C. Modeling and evaluation . ２９
D. Automated model maintenance . ３１
2.7. Real-world deployment ３３
A. Generality . ３４
B. Ground truth and model maintenance ３５
C. Extreme cases and anomalies ３５
D. Hacks detection ３７
E. False positive issue ３７
F. Individuals using bots ３７
2.8. Conclusion ３８
Chapter 3. Gold farming group detection using network analysis ４０
3.1. Introduction . ４０
3.2. Related works . ４２
3.3. How RMT works ４３
3.4. Dataset . ４４
3.5. Trade network analysis for RMT detection . ４５
A. Construction of a virtual goods trading network and community detection ４５
B. Community grouping by network structure ４７
C. User grouping by play style ４９
D. RMT group detection . ５１
E. RMT estimation . ５５
3.6. Measurement study of RMT ５７
A. Volume of RMT . ５７
B. Monopoly of RMT market ５９
C. RMT network in other games ６１
3.7. Conclusion ６２
Chapter 4. Profit optimizing churn analysis ６４
4.1. Introduction . ６４
4.2. Related works . ６６
4.3. Dataset . ６８
A. Time unit for data aggregation . ６９
B. Churn definition ６９
4.4. Methodology ７０
A. Prediction target selection . ７０
B. Feature engineering ７５
C. Profit evaluation ７８
4.5. Exploratory data analysis ８１
A. Trends and variations in gaming activities . ８１
B. Party activity . ８２
C. Legion ８４
4.6. Experiments . ８６
A. Total churner prediction vs. loyal churner prediction . ８６
B. Threshold optimization ８９
4.7. Discussions . ９０
A. Limitation of a binary classifier . ９０
B. Cost optimization . ９２
4.8. Conclusion ９３
Chapter 5. Limitations, alternatives and future works ９５
5.1. Limitations ９５
5.2. Alternatives ９６
A. Weak supervision . ９６
B. Feature embedding . ９６
C. Expected profit function ９９
5.3. Future works １００
Chapter 6. Conclusion . １０１
References . １０３
Acknowledgements . １１２
