데이터 분포변화와 머신러닝 성능 변화 관계 분석을 통한 재학습 시점 판단 프레임워크

초록

The adoption of AI-based systems has rapidly expanded across various industrial domains. However, the performance of deployedmachine learning models can significantly degrade when the distribution of incoming data changes over time. Although prior studieshave focused on monitoring data distribution and model performance changes, relatively limited research has systematically investigatedtheir relationship to determine appropriate retraining timing. This paper proposes an AI model retraining decision framework thatquantitatively analyzes the relationship between data distribution changes and predictive performance to identify effective retraining points. To validate the framework, two case studies were conducted using airport environmental data and stock market data. Experimental resultsshow that the proposed framework reduced retraining frequency by 84.2% and 81.4%, respectively, compared with annual retraining,while maintaining comparable performance for weather data and improving performance for financial data. These results demonstratethat the proposed framework can effectively determine retraining timing and reduce retraining costs.

키워드