| Summary: | 碩士 === 國立政治大學 === 資訊管理學系 === 107 === Customer churn has long been recognized as one of the most important predictive issues. Through customer churn prediction, companies can know the likelihood of a customer repurchasing in the future, as well as the exact timing of the repurchase. In the past, there have been many different areas exploring the repurchase predictions, and these areas have developed individual prediction models. However, the lack of discussions and comparisons of models in different areas motives the research. This study proposes a hybrid model that integrates the BG/BB (Beta Geometric/Beta Bernoulli) probability model frequently used in the field of marketing, and the neural network commonly used in computer science and artificial intelligence. The hybrid model using transaction data of current season and the BG/BB parameters has improved the prediction performance (average precision) by 6.5% and 8.1% compared to the two individual models respectively. The improvement indicates that the statistical marketing model and the neural network can complement to each other. We further perform data clustering and identify the set of customers with better predictability. Especially, we find that using Recency instead of K-Means as the clustering indicator has lower computational costs and more interpretability while the prediction performance is similar. We also compare the hybrid model and the LSTM model. The finding indicates that the hybrid model has better prediction performance than complex time series model, and the cost of modeling is even lower. The study has both practical and academic contributions. First, the proposed hybrid model can help companies improve forecasting accuracy with relatively low cost, thereby improving the return on investment of marketing. In addition, while the development of repurchase forecasting in marketing and data mining has been very vigorous, this study is the first to integrate models across areas and presents the process of building the hybrid model and further evaluate its performance.
|
|---|
