Modeling Sub-Document Attention Using Viewport Time Max Grusky Jeiran Jahani Josh Schwartz Dan Valente Yoav Artzi Mor Naaman (with support from Nir Grinberg)
Current understanding of user engagement on the Web is limited mostly to the document level . bounce rate page views time-on-page - OVERALL GOAL Understand how users interact with documents online. - IN THIS WORK Design and validate a method of measuring sub-document user attention.
------------------------------------------------------------------------------------------------------------------------------------------------------------------ OUR APPROACH - Develop a model of sub-document attention by 1 building on results of prior small-scale lab studies. Validate our sub-document attention model on 2 large-scale Web data using a known user behavior metric . 1.2 million reading sessions Cross-language on a popular news site. reading rate.
Eye tracking to measure attention + Fine-grained measurement of attention + Great for understanding engagement patterns – Expensive, dedicated hardware – Calibration and lab setting – Does not scale to millions of web users Our approach: measure attention using only standard browser data .
ESTIMATED USER ATTENTION Uniform attention – Divide total user attention uniformly across the page, based on page element size. – Does not take into account viewport information. ‣ What can we learn about attention from lab studies?
(Sharmin et al., 2013) Viewport attention distribution is very predictable! (Buscher et al., 2010)
ESTIMATED USER ATTENTION Gaussian viewport attention – Uses the user’s viewport and attention distribution to assign attention to page elements. + Empirically motivated: Builds on prior research in attention. ‣ How do we validate it?
User engagement dataset Cross-language reading rate – 1.2 million reading sessions across a – Readers have predictable and popular new website collected by measurable rates across nine languages. Chartbeat, Inc. (Susanne Trauzettel-Klosinski and – Each session consists of second-to- Klaus Dietz, 2012) second viewport time data. 5 sec. → 240 WPM Validation Approach: Reading rates estimated 12 sec. → 220 WPM by the attention model should correspond with known reading rates for each language. 8 sec. → 100 WPM
Uniform A)en+on Model Uniform A)en+on Model Gaussian Viewport A/en0on Model 250 English English 600 600 600 Es#mated (WPM) Es#mated (WPM) Es#mated (WPM) English Spanish Spanish 200 400 400 400 German German Spanish German 150 200 200 200 200 200 200 400 400 400 600 600 600 150 200 250 Empirical (WPM) Empirical (WPM) Empirical (WPM) Number of Readers Number of Readers Number of Readers 4000 4000 4000 8000 8000 8000 12000 12000 12000 16000 16000 16000
Image Elements Image Elements War War US Poli'cs US Poli'cs Applying our model UK Poli'cs UK Poli'cs What about images? Which article topics Sex Sex use the most engaging images? Police Police Music Music – Group articles by topic. Health Health Gaming Gaming – Apply our viewport-based sub-document attention model to each session. Fashion Fashion Drugs Drugs 0 sec. 0 sec. 10 sec. 10 sec. 20 sec. 20 sec. 30 sec. 30 sec. Average Dwell Time Average Dwell Time
------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------ IMPLICATIONS & APPLICATIONS CONCLUSIONS - - Understanding language Attention measurement at scale – Write automatic summaries using sub- – We can reliably measure sub-document document attention distributions. attention within the browser. – Especially useful tool at scale. Understanding engagement – Such as measuring attention of millions – Help identify and aid struggling readers. of users to thousands of images. – Better understand user preferences. Max Grusky Modeling Sub-Document Cornell University Attention Using Viewport Time grusky@cs.cornell.edu
Recommend
More recommend
