|
Article Excerpt
Innovative methods of artificial intelligence such as artificial
neural networks (ANNs) have been increasingly adopted to predict consumer responses to direct marketing. However, appropriate learning algorithms, evaluation criteria, and validation procedures are necessary for effective implementation of neural networks to provide decision support to managers. This study compares the performance of Bayesian neural networks with that of logistic regression and the backpropagation method in modelling consumer responses. The results of a tenfold stratified cross-validation suggest that although the three methods perform equally well under the error rate, Bayesian neural networks generate higher statistics for the Area under the Receiver Operating Characteristic Curve (AUROC) and cumulative lifts. The findings suggest that researchers should adopt effective learning algorithms, relevant evaluation criteria and appropriate validation procedures for neural networks to model consumer responses and solve marketing problems facing today's businesses.
Introduction
Modelling consumer responses is critical for direct marketing operations to increase sales, reduce cost and augment profitability. Researchers have developed various methods to model consumer responses to direct marketing. A common approach is to use logistic regression to classify consumers based on their probability to purchase from a specific promotion (Berger & Magliozzi 1992). Given the tremendous growth of direct marketing in recent years, accurate prediction of consumer responses to direct marketing has become a top priority for many companies (Bodenberg & Roberts 1990). Meanwhile, the rapid accumulation of customer and transactional data has prompted marketers to search for innovative data-mining methods to extract useful knowledge that can assist decision-making (Peacock 1998; Shaw et al. 2001). Recently several researchers have applied artificial neural networks (ANNs) to model direct marketing responses and their efforts have produced encouraging results (Zahavi & Levin 1997b). However, ANNs learned with the backpropagation method are known to suffer from over-fitting the training data and result in poor performance with real marketing data. Recently, the Bayesian approach to learning neural networks has been proposed to minimise over-fitting and it has proved to be a viable alternative to modelling consumer responses to direct marketing (Baesens et al. 2002).
In addition to a feasible learning algorithm, evaluation criteria and validation procedures are necessary for effective implementation of neural networks in direct marketing (Ratner 2003). First, in modelling consumer responses to direct marketing, the cost of misclassification is unequal for different types of errors and in fact higher in top deciles than in the lower ones. Thus, relevant evaluation criteria are critical for comparing the performance of alternative methods and, more importantly, to provide better decision support for managers (Rosset et al. 2001). Second, due to the low response rate, direct marketing data often have a small proportion of buyers (positive cases). How to validate the results of neural networks presents a significant challenge. In this study, we compare multiple evaluation criteria, including error rate, statistics for the Area under the Receiver Operating Characteristic Curve (AUROC), and cumulative lifts. To validate the neural network models, we adopt the tenfold stratified cross-validation. The results of experiments with a direct marketing dataset indicate that although three methods perform equally well under the error rate, Bayesian neural networks generate higher statistics for the AUROC and cumulative lifts than logistic regression and the backpropagation neural networks. These results suggest that researchers should adopt effective algorithms, relevant evaluation criteria and appropriate validation procedures for neural networks to model consumer responses and solve marketing problems facing today's businesses.
Models of consumer responses to direct marketing
Direct marketing has witnessed phenomenal growth worldwide in the past decade (Bodenberg & Roberts 1990; Bult & Wansbeek 1995). According to the statistics of the Direct Marketing Association (2002), consumer sales generated through direct marketing channels in 31 countries reached US$2.28 trillion in 2000 and will grow at an annual rate of 13% until 2005, accounting for a significant portion of the economic activity in these countries. The expenditure of direct marketing to consumers will grow roughly 9.8% annually during the same period. In the UK alone, the total expenditure on direct marketing has grown by 5.9% since 2000 to reach a value of [pounds sterling]18 billion or US$26.7 billion in 2001 (Euromonitor International 2002). The total value of the market has increased by 35% over the same period in the UK. Besides the traditional direct marketers such as catalogue companies and telemarketers, many large corporations have adopted direct marketing as one of their strategies. Accurate prediction of consumer responses to direct marketing has become increasingly important for these businesses.
Conventional methods
Modelling consumer responses to direct marketing using conventional research methods has largely been based on the a priori approach. For instance, direct marketing researchers have developed a theory of consumer response, known as the RFM model. The model states that the likelihood of consumers responding to a direct marketing promotion is predicted by the recency of the last purchase, the frequency of purchase over the past years, and the monetary value of a customer's purchase history (Berger & Magliozzi 1992). A better model is said to correctly classify a greater percentage of the responders at a given depth of file. Based on the results of the model, consumers are divided into different segments based on their response probabilities. More complicated models may include consumer demographic and psychographic variables, credit histories and purchase patterns (Berry & Linoff 2000).
Researchers to date have adopted various methods to predict consumer responses to direct marketing, such as logistic regression, discriminant analysis and multiple regressions. Various models of consumer response have emerged in the literature. For instance, DeSarbo and Ramaswamy (1994) have proposed the Customer Response-Based Iterative Segmentation Procedures (CRISP), which can simultaneously derive market segments and estimate models of customer response in each segment. By controlling for unobserved heterogeneity among consumers, this model can help to improve the accuracy of classification. Several authors have tested a beta-logistic model that could update the estimated response probabilities over time and lead to more accurate predictions (Rao & Steckel 1995). Tree-generating techniques, such as CHAID and CART, have also been used to predict consumer purchase (Haughton & Oulabi 1997). Improving the accuracy in predicting consumer responses to direct marketing is the main objective of these methods.
Despite the progress, consumer responses to direct marketing have remained low--for instance, 2-4% or even less. One of the reasons for such dismal results is the violation of the key assumptions of the research methods when the models are applied to real data. In addition, conventional statistical approaches have several limitations. First, they can typically handle only a limited number of variables, which are subject to a number of assumptions and constraints on the types of data and their distribution. Second, the traditional methods are largely based on fixed-form equations such as logistic regression and treat consumer response as a linear additive model. Studies using linear models assume a single best solution and can compare only a few alternative solutions manually....
|
