Response modeling in the domain of customer relationship management (CRM) is to predict certain customers who are more likely to buy a promoted product or service based on the customer database. The customers who have been targeted in a preliminary campaign have the labels of either respondent or non-respondent and thus are called labeled, whereas the others, who have been excluded from the campaign and thus do not have labels, are called unlabeled. Because of the limited marketing cost for the campaign, only a small part of the customer data is labeled whereas the rest large amount of the customer data remains unlabeled. In the light of that, the most recently proposed semi-supervised learning (SSL) is well-suited for response modeling. In the framework of SSL, the labeled data points are used for the basic modeling and the unlabeled data points are also used for exploiting the information on the underlying manifold structure of input space. However, there are some technical difficulties. Because of the lack of the labeled data points, it becomes difficult to find the optimal value of the learning hyperparameter and to figure out the influence of noise on performance. To circumvent the addressed difficulties, we propose to employ ensemble learning and graph sharpening. The ensemble learning replaces the hyperparamer selection procedure to an ensemble network of the committee trained with various values of hyperparameter. On the other hand, graph sharpening helps to remove unhelpful information caused by noise. The proposed method was evaluated on an artificial data and benchmark data. And we applied to the real-world response modeling for the insurance promotion provided by the CoIL Challenge 2000. The experimental results present the proposed method enables response modeling with only a few customer data without concerning the technical difficulties, selecting the best value of the hyperparameter and mitigating the influence of noise.
