A personalized content-based method to predict customers’ preferences in an online apparel retailer

A critical decision for an online retailer is to select a set of products out of thousands of possible choices to present to the customers on a web page. The retailer may prefer to offer a different set for each customer because customers have heterogeneous preferences. Thus, to offer the optimal set, the retailer needs to know the customer’s preferences. We propose a new personalized content-based method to comprehend customers’ preferences in an online retailer based on customers’ previous clicks and purchases and attributes of the products. We represent each product with an attribute vector that consists of all attributes of a product, e.g. color and brand. Then, for each customer, a score is assigned to each attribute vector based on the customer’s previous preferences, representing his/her interest in that combination of attributes. We test the method using data provided by an apparel retailer. Our method outperforms benchmark methods (including Collaborative Filtering) in predicting customers’ preferences (i.e. clicks and purchases) in general, and it has a strictly better performance in predicting customers’ preferences over new products. Also, our method outperforms benchmark methods with a high margin in predicting the preferences of customers who are not generally interested in popular products. Finally, we implement a hybrid method consisting of all implemented methods named the Smart Selection. This method outperforms all methods in predicting clicks and purchases with a high margin. This shows that our method provides a complementary approach for Collaborative Filtering by successfully addressing the limitations of commonly used methods.