ClickGraph: Web page embedding using clickstream data for multitask learning

Abstract

The rise of big data frameworks has given website administrators the ability to track user clickstream data with more detail than ever before. These clickstreams can represent the user's intent and purpose in visiting the site. While existing work has explored methods for predicting future user actions, these methods are limited focus solely on one task at a time, ignore graph structure inherent in clickstreams, or model the conversion of the entire clickstream session, ignoring complexities such as multiple conversions in a single session. In this work, we formulate the novel problem of simultaneously predicting future user actions given a user's clickstream history. We argue that clickstream data contains important signal for predicting future user action. To tackle this new problem, we propose a novel method called ClickGraph, a recurrent neural network that encodes the graph structure of user click trajectories in the learned representations of web pages. We conduct experiments on a real-world dataset and demonstrate that this multitask learning approach is effective at improving the prediction of form fill conversions over strong baselines. In particular, we demonstrate that the ClickGraph model is effective at reducing false positive rates, increasing F1 scores, and improving recall.