Recommender system for marketing optimization

Most of existing e-commerce recommender systems have been designed to recommend the right products to users, based on the history of previous users’ individual transaction records. The real application scenarios of recommendation also have different requirements. From the customer point of view, many users visit the websites anonymously, so a practical way to provide anonymous recommendation is needed. From the marketing point of view, the recommendation list is not only a place to display the correlation of products, but also a place to display the variety of products as well as a tool to promote products. From the data point of view, concentration bias may be a serious problem. In this paper we propose trigger and triggered (TT) model to address all of these issues. First, the proposed model generates trigger and triggered pairs with significant correlations which can be used either to create a practical anonymous recommendation or as an input for products lifecycle modeling. The generated pairs not only reflect the relationships between products but also solve the problem of concentration bias very well. Besides, exposure of products required by marketing can be accomplished in the modeling. Second, by using the pairwise knowledge from the first step, the proposed model can recommend the right product at the right time to stimulate future consumptions and increase customers’ engagement for the off-site case. A real-life retail store data is used to evaluate the proposed model, and the experimental results show that the model can decrease the problem of concentration bias while improving the correlation between recommendation items. The TT model significantly improves the sequential purchases on triggered items.