Recommendation Systems for Sequential Decisions

Although the marketing literature on recommendation systems has demonstrated the value of incorporating consumer behavior models to improve algorithm performance, most studies have focused on situations in which choices are simultaneously available, such as recommendations on Amazon. In this paper, we propose a recommendation system for sequential decision-making scenarios in which a series of choices are made sequentially available and expire quickly, and the quality of future choices is uncertain. Recommendations on daily deal websites provide an example of such a scenario. The proposed recommendation system is grounded in the search literature of consumer behavior. We develop a new machine learning algorithm to train the model and make predictions. Specifically, we leverage the means of semiparametric Gaussian copulas to generate a complex joint distribution among variables, including individual characteristics, deal features, a sequential temporal factor, and purchase decisions. Recommendations are made based on each user’s purchase likelihood for each deal available on a specific day, calculated by the conditional posterior of the joint distribution. The algorithm’s semiparametric nature and maximum likelihood estimation method make it robust to imbalanced data. We apply the model to a proprietary data set comprising Groupon customers’ clickstream data and demonstrate the superior performance of the proposed approach by comparing it with other popular recommendation algorithms for sequential decision-making scenarios. The proposed system facilitates companies’ efforts to target the right customers with the right product at the right time.