Using Bayesian network for purchase behavior prediction from RFID data

This paper represents our recent studies about the prediction of purchase behavior and an advancement of in-store behavior with respect to RFID technology. In contrast to prior innovators in this research field, this paper has paid special attention to stay time spent on shopping in a target area rather than the whole supermarket, which can serve us to interpret the decision process of purchasing one product or a series of products in a much more intuitive and precise measurement. Also, we develop an integrated model to combine purchase behavior and in-store behavior. A probabilistic graphical model - bayesian network is employed to demonstrate a quantitative analysis process of purchase behavior decision over stay time. In order to distinguish purchase intention among different customers, an attitudinal factor - purchase background of customer is introduced in this paper to build bayesian network. As bayesian network can only deal with the discrete variables, a clustering algorithm is applied to discretize the continuous variables. In the experiments, the optimal cluster number of stay time and purchase background is examined for maximizing the performance evaluation with higher accuracy, and the results also show bayesian network has a better accuracy than other typical prediction models. Finally, we investigate the sensitivity and specificity of purchase behavior predicted by our proposal in adjustment of decision threshold, and use ROC (Receiver Operating Characteristic) curve to determine the optimal decision threshold which can maximize the classification accuracy of models.