Modern marketing applies data analytics to determine measures that influence a customer’s future purchase decisions based on past behavior. One aspect of this behavior is that customers like to vary their purchases over time by choosing alternative products or shops. This paper aims to improve our understanding of variety-seeking behavior via a data analytics framework that determines the important features influencing purchase decisions and predicts the likelihood of future variation.

The methodology is not just based on data about past consumption, but also demographic data (from loyalty cards) and environmental data (time of purchase and current weather conditions). From this, the preprocessing step selects an appropriate feature set and prepares a dataset that is suitable for machine learning. Part of the dataset is used to train four standard machine learning algorithms, whose results are then combined by an ensemble learning method to a joint model.

The experiments performed on a large dataset of Chinese university students who buy meals in several canteens with their student cards (apparently data privacy considerations do not play a role) are particularly interesting. The results demonstrate that past consumption and environment are the most important features, and considering a past consumption record of only seven days gives the best results. While the achieved 67 percent precision for behavior prediction is not overly impressive, more features will be investigated to yield more accurate models.