Changing perspectives: Using graph metrics to predict purchase probabilities

Abstract

The prediction of online user behavior (next clicks, repeat visits, purchases, etc.) is a well-studied subject in research. Prediction models typically rely on clickstream data that is captured during the visit of a website and embodies user agent-, path-, time- and basket-related information. The aim of this paper is to propose an alternative approach to extract auxiliary information from the website navigation graph of individual users and to test the predictive power of this information. Using two real-world large datasets of online retailers, we develop an approach to construct within-session graphs from clickstream data and demonstrate the relevance of corresponding graph metrics to predict purchases.