Thesis Proposal - Conor Kelton - Towards a More User-Centered Web Experience

Thursday, May 21, 2020 - 12:00pm to 1:00pm
Event Description: 

Conor Kelton presents his thesis proposal, "Towards a more User-Centered Web Experience"
Studies from major players in the Web ecosystem, such as Amazon, Walmart and Google, show that small improvements to Web page load performance can have an enormous downstream impact on revenue and end-user satisfaction. However, we posit common metrics used to
quantify Web loads fail to account for the user perception. As users are unlikely to interact with the page until they perceive it to be loaded, we design a metric to capture when such a time occurs, called the user's perceived page load time. We use crowd-sourcing to collect the users’ perceived page load time across 450 desktop and mobile users as they view over 100 Web pages. For desktop browses, we asked users to respond with a simple mouse click when they thought the page load was complete. On mobile browsers, the important content is spread across multiple viewports, causing users to scroll during the page load. So, we create a specialized user study to collect ground truth user perceived load time for mobiles while allowing users to scroll.

Our key finding is that across desktop and mobile browsers, metrics used to measure page load times do not correlate with the user-perceived page load times. As bettering the user perception is critical, we next explore how to modify pages to actually reduce the user's perceived page load time. To do so, we leverage gaze tracking, which is commonly employed in the cognitive science literature to measure a user's focused attention. Our gaze study shows that certain objects on the page are more important than others based on the collective fixation across users. We develop WebGaze, a system which prioritizes bandwidth to objects which demand user attention over those that do not during page loads. We show WebGaze improves the user perceived page load over state-of-the-art optimizations with crowd-sourced studies of a new set of 50 users across 50 pages.

Measuring the user perception using large-scale crowd-sourcing is not practical for most Web developers. We introduce a model that can predict the user's perception of page loads. We design the model to be as accessible to Web developers. As such, the only inputs required to
use the model are standardized Web page load metrics developers readily have access to. The ground truth used to train this model is the user's perceived page load time. The result is a trained model that can predict when users perceive pages to be loaded across device types. We show how this model can be used to help developers evaluate the effects of their (mobile) Web page optimizations.

One problem with measuring the ground-truth user perception of page loads is the need to run in-lab user studies to obtain these measurements. We propose instead to measure the user's perceived page load time implicitly by finding when users engage with pages using only their eye gaze. To do so, we look for when repeated long pauses occur the user's fixation behavior, and mark this as the time when they actually engage with important content on the page. We call this measure the engagement onset time for the user, and propose to use this time to engagement to quantify the user experience with pages. Commodity webcam trackers can be used to obtain this engagement behavior in the wild without the need for explicit in-lab studies. For our proposed work, we will show the feasibility of using such trackers for our purposes. We will then study how the onset time changes when adjusting the load order, or even withholding, objects during the page load. We envision our experiments can help developers understand how to change the load process of their pages to minimize the time users need to become engaged and their perception of when pages are loaded.

For Zoom information, please email

Computed Event Type: 
Event Title: 
Thesis Proposal - Conor Kelton - Towards a More User-Centered Web Experience