[Colloquium] Yiyang Ou MS Presentation/May 28, 2021

[Jessica Garza]

Yiyang Ou is a student in the Bx/MS program.

—————————————————————————————————————————————

Date: Friday, May 28, 2021

Time: 10 AM, CST

Location: remote via Zoom <https://uchicago.zoom.us/j/5711251740?pwd=TSsxVCt1bFdhRjcrY2FmeDEveHY1QT09 <https://uchicago.zoom.us/j/5711251740?pwd=TSsxVCt1bFdhRjcrY2FmeDEveHY1QT09>>

Meeting ID: 571 125 1740
Passcode: 166500

M.S. Candidate: Tony Ou

M.S. Paper Title: Improving web QoE metric with saliency

Advisor: Prof. Junchen Jiang

Committee Members: Prof. Junchen Jiang, Prof. Shan Lu, and Dr. Siddhartha Sen


Abstract:
—————————————————————————————————————————————

QoE (Quality of Experience) metrics for web page loading are essential for web developers to quantify web pages' performance and improve users' experience. However, traditional metrics, such as Speed Index, are known to have low correlations with user rating. One weakness is that they are agnostic to or only crudely accounting for user attention, such as the Above-the-Fold (ATF) time metric classifies the web objects into two groups: above-the-fold and below-the-fold. The downside of this is that they miss the optimization opportunity in changing the loading order of web objects to match user attention patterns. For example, we can load objects users care about first and delay the objects users ignore to improve users experience without shortening the overall load time.

In our experiment, we incorporate two attention patterns to improve traditional web QoE metrics. First, we give more reward to web pages for loading objects close to the center of users’ attention and objects likely beyond users’ attention. Secondly, we penalize the web pages for visual changes that occur late in the loading process and attract user attention, since they make people think the page is not loaded. However, it is not an easy task to predict user attention. Past work has used actual human study to track human gazes, but this is not scalable for practical usage. Instead, we utilize recent computer vision works on saliency models to predict user attention. We built an end-to-end pipeline that, given videos of webpage loading, it computed saliency augmented speed index that captures the impact of user attention, We evaluated its performance on a web load dataset with human ratings, and saw improvement in the correlation of our metrics with human scores over that of traditional metrics.

—————————————————————————————————————————————



Jessica Garza
Assistant Director of Undergraduate Studies
Department of Computer Science
The University of Chicago
Covid-19 Resources <https://cs.uchicago.edu/remote2020/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20210525/2e420ee2/attachment.html>