Publication Details

Download

Changkun Ou
Understanding and Predicting Web Browsing Behavior
Master's Thesis, February 2019. (bib)

Clickstream applications appeared at the end of last century and have proliferated the heart of our Internet world. Trades, public opinions, and almost every web requests are precisely recorded on server-side log files. The fundamental interaction between a web service client and server stands immutably, even though mobile devices have governed our daily life. This thesis proposes a machine learning model that characterizes user browsing behavior while including multi-tab branching and backtrack actions in a browser instead of web request-based clickstreams. The model is named the Action Path model (APM). To justify the APM, a lab study is established and individuals' clickstream data is collected, which consisted of chronologic URLs and corresponding stay durations for each URL. The thesis designed nine different contexts given web browsing tasks for three mainstream websites based on the theory of information behavior. Each website has three types of tasks: a goal-oriented task, fuzzy task and exploring browsing task. They characterize the corresponding three browsing behaviors. The thesis seeks to achieve the following goals by analyzing the subject's trace from a lab study: 1) Understanding: identify if browsing behaviors are distinguishable and find common patterns that appear in an action path. 2) Classification: separate and report browsing behaviors on the web, which will help users to better understand their status. 3) Prediction: present the future click path in more than one step with the given context of the browsing history in a session. The quantitative analysis in this thesis indicates that goal-oriented, fuzzy, and exploring browsing behaviors are classifiable with 100\% accuracy based on the combination of chronologic URLs and stay duration. The prediction performance of APM indicates higher than 60\% accuracy for three to five steps of future clickstream prediction. The qualitative analysis of the APM indicates five observed patterns, including ``ring'',``star'', ``overlap'', ``hesitation'' and ``cluster'' patterns, which represent the patterns of an action path. To illustrate the application, a browser plugin is developed that proactively serves users, and suggests predictions for the possible future user clicks. Furthermore, the thesis discusses a generalized design of APM and plugin communication protocol. This discussion explores the possibility of formalizing the model and protocol as standard Web APIs to help designers and developers to improve and monitor the user experience of their products. To the best knowledge of the author of the thesis, the proposed APM is the first model with a detailed study regarding web browsing behavior modeling based on clickstreams collected from the client side.