Algorithmic Recommender Systems for Personalized E-Commerce

Abstract

Although many personalisation systems focus on predictive accuracy, they miss the point of personalisation in e-commerce, which is based on real-time responsiveness. This study suggests and empirically validates a real-time recommendation system based on a hybrid collaborative and content-based scoring function which uses customers' clicks, views of the products and the products they have purchased in the past to provide personalized product suggestions, along with a dedicated latency optimization layer. A stratified random sample of 200 platform users was divided into three strata based on their activity on the platform and randomly allocated to a Treatment condition (recommendation engine shown, n = 97) and a Control condition (non-personalized listing, n = 103). The Treatment group also had significantly higher rates of click-throughs, conversion rate, satisfaction, perceived relevance, perceived speed, and likelihood to return, with significantly lower response latency, than the Control group (all p < .001, Cohen's d = 1.27-3.19). Multiple regression analysis also revealed that Precision@10, Recall@10, NDCG@10 and response latency together account for 75.1% of the variance in the sales uplift, accuracy and response latency are both significant and independent with one another. The results in this report prove that predictive accuracy and system speed do not need to be at odds, and that a properly designed hybrid recommendation system can improve the customer experience and business results. The study provides an empirically grounded mathematical and software approach to improve the personalization of e-commerce while maintaining real-time system performance that e-commerce practitioners can adopt.