데이터 특성에 따른 AI 모델의 고객 이탈 예측 성능 비교 연구 = A Comparative Analysis of AI Models’ Performance in Predicting Customer Churn Across Different Data Characteristics

다국어 초록 (Multilingual Abstract)

Customer churn prediction is a critical activity necessary for startups to continuously retain their acquired customers. Extant research has proposed techniques for predicting customer churn by analyzing unstructured data using artificial intelligence technologies. However, these studies have limitations in that they lack consideration of the linguistic features embedded in the data and rely heavily on publicly available datasets. Therefore, our study collected 2,616 Voice of Customer (VoC) data from a domestic Korean telecommunications company (Company A) and generated three datasets based on linguistic characteristics: (i) using multi-turn conversation text alone, (ii) using multi-turn conversation text combined with sentiment scores, and (iii) using multi-turn conversation text combined with stylistic features. We performed classification analysis on these three datasets using sequence AI models (LSTM, Bi-LSTM, GRU) to distinguish between churned and non-churned customers. The analysis results revealed that the linguistic characteristics embedded in the datasets are more important for customer churn prediction than the AI models themselves, such as LSTM, Bi-LSTM, and GRU. These findings provide methodological implications for prior research that has focused solely on improving analytical model performance and offer practical guidelines for startups that need to predict customer churn.