Identifying Price Sensitive Customers in E-commerce Platforms for Recommender Systems

Abstract

With the rise of E-commerce platforms, more people are getting used to make their daily purchases in online stores, especially for price-discounted goods. Therefore, online price cutting campaigns become a common approach for online retailers to compete with other competitors. Still, giving the same price discount to all may not be an efficient resource allocation as different users respond differently due to the differences of their price sensitivities. So if we are able to identify price sensitive users, both sellers and recommender systems will be greatly benefited from it in terms of improved user targeting and item suggestions. However, due to lack of detailed historical price and customer profile data, it is challenging to conduct price sensitivity analysis via traditional economics approach. More importantly, it is really hard and costly for companies to acquire price sensitivity labeled data. To overcome the constraints, making use of rich meta data (e.g. comment reviews) and time stamp becomes an alternative way. Inspired by distinct expressive power of graphical model, especially bipartite graph, we propose a User Behaviour Probability Transition Model (UBPT) which considers both user and item price sensitivities as weightings in the probability transition process. First, we define our own set of price sensitive users according to anonymous user after-purchase reviews. Second, we integrate selected behavioral features via doing user and item encoding. Third, using both user and item similarities, we combine our algorithm to simulate the probability transition process. With the data set from JD.com, our proposed model significantly outperforms other baselines in most cases. Besides, through applying the idea of UBPT to recommender systems, we can also enhance the performance of traditional recommendation algorithms.