CRM决策的序贯优化——以KDD98直邮决策为例

Sequential Decision Optimizing in CRM——An Example of KDD98

从多阶段、延迟回报的角度来看待CRM中的决策优化问题。以KDD98数据集为例,将邮寄序贯决策定义为一个部分可观察马尔可夫决策模型(POMDP)。提出了模型参数估计的EM算法并用MATLAB实现;用模型对数似然值、BIC统计量选择最佳模型;用向前一步预测对模型进行检验;用Incremental prune算法对模型求解。实证结果表明,POMDP模型可以很好的捕捉客户购买行为的动态变化,对客户的购买有很好的预测效果。在此基础上,说明了如何使用该模型以客户终生价值最大化为目标优化直邮策略。

We consider the problem of CRM decision optimizing from a multistage and delayed reward perspective. Taking the well-known KDD98 dataset as an example, we describe the sequential mailing decision with a partial observable Markov decision process. We propose an EM algorithm for model parameters estimation, and implement the algorithm in MATLAB. We use Log-likelihood statistic and BIC for model selection, one-step-ahead prediction for model validation, and incremental prune algorithm for model solution. The results show that the proposed POMDP model captures customer＇s dynamic donating behavior very well, and can predict the purchasing probability accurately. Based on the model validation, we further show how to solve the model to get the optimized mailing policies while maximizing the Customer Lifetime Value.