The research on Optimal media selection models Using Data Envelopment Analysis and neural networks

从数据包络分析和神经网络的角度对媒体选择模型的研究进行了综述。我们查阅了一些学术期刊，如《管理科学》、《市场科学》、《市场研究杂志》、《市场营销杂志》等。同时，我们也对数据包络分析和神经网络本身进行了调查，以跟踪它们的发展。许多广告活动数据是从几个来源收集的，为了分析我们的研究，它们被重组了。经过一些筛选过程，61个广告活动数据可供我们分析。采用基于随机前沿的数据包络分析模型建立了最优媒介选择模型。我们开发的模型的优点是不需要对功能形式进行估计。利用蒙特卡罗模拟生成的人工数据，对既有隐式Cobb-Douglas和凹边函数的问题，也有隐式非Cobb-Douglas和凹边函数的问题进行回归分析，发现该模型比基于参数化方法的结果更好。通过将我们的模型应用于广告活动数据，我们得到了最优解。在日本市场科学研究所的会议上做了报告后，一些广告公司的成员对我们的方法给予了很高的评价。对于神经网络，我们搜索了许多软件包软件。我们可以得到一些，SAS研究所的Enterprise Miner， SPSS的Clementine， S-PLUS和Predict of Neuralware。利用样本数据考察了各包软件的预测能力。其中，从使用简单和输出准确的角度来看，很明显，Predict of Neuralware优于其他三款软件。

We surveyed the works concerning about media selection models from the point of data envelopment analysis and neural networks. we checked some academic journals like Management Science, Marketing Science, Journal of Marketing research, Journal of Marketing etc. At the same time we also surveyed data envelopment analysis and neural networks themselves so as to follow their development.Many advertising Campaign data were collected from several sources and they are restructured in order to be analysed for our research. After some screening processes, 61 advertising campaign data were available for our analysis.We developed the optimal media selection model using data envelopment analysis model based on Stochastic frontier. The merit of model that we developed does not need the estimation of functional form. Using the artificial data we generate by monte-carlo simulation, we found our model can obtain better result than parametric method based on the regression analysis in the problem with implicitly not only Cobb-Douglas and concave frontier function but also non-Cobb-Douglas and concave frontier function.By applying our model to advertising campaign data, we got the optimal solution. After the presentation at the conference of The Japan Institute of Marketing Science, some member of advertising agency set a high valuation on our approach.As for neural network, we searched many package soft for neural network. we were able to get some, the Enterprise Miner of SAS Institute, Clementine of SPSS, S-PLUS and the Predict of Neuralware.Using sample data we investigate the prediction ability of each package soft. Among them, it become clear that the Predict of Neuralware outperform the other three softs from the point of the simplicity to use and the accurate output from it.