We test if modern computer-vision algorithms can predict if users are reading relevant information, from their eye movement patterns. The slides accompany the video presentation at https://youtu.be/ZebBgUhL-EU
The full research paper is available at:
https://dl.acm.org/doi/10.1145/3343413.3377960
and also at
https://arxiv.org/abs/2001.05152
2. Nilavra Bhattacharya1, Somnath Rakshit1, Jacek Gwizdka1, Paul Kogut2
ACM SIGIR CHIIR 2020 • VANCOUVER VIRTUAL
RELEVANCE PREDICTION
FROM EYE MOVEMENTS
Using Semi-interpretable Convolutional Neural Networks
1 School of Information, The University of Texas at Austin
2 Rotary and Mission Systems, Lockheed Martin Corporation
ixlab.ischool.utexas.edu
4. Two Worlds in the Information Field
Introduction User Study Scanpath Image Classification Interpretability Conclusion
Image: Tefko Saracevic (https://studylib.net/doc/15399702)
• Situational relevance or utility:
“situationally relevant items of
information are those that
answer, or logically help to
answer, questions of concern”
(Wilson, 1973)
• This work: situational relevance
= users’ perceived-relevance of
the documents they examine for
answering a question
Image: https://www.noldus.com/applications/eye-tracking-
physiology
Eye-tracking
5. Background: Eye-tracking & Information Relevance
Introduction User Study Scanpath Image Classification Interpretability Conclusion
• Drawback 1: aggregate ET data at stimulus / trial / participant level
• aggregated fixation counts/durations (Fahey+ 2011; Frey+ 2013; Gwizdka 2014; Loboda+ 2011; Puolamäki+ 2008; Wenzel+ 2017; Wittek+ 2016)
• reading related preprocessing before aggregation (Buscher+ 2008; 2012; Gwizdka, 2014a; 2014b, 2017; Gwizdka+ 2017)
• ET features from 2-second windows near the end of trial has more discriminating power (Gwizdka+ 2017)
=> collapsing ET data leads to loss of information
• Drawback 2: lack of standard feature selection => varied prediction performance; accuracy rarely above 70%
(Simola+ 2008; Slanzi+ 2017; Wenzel+ 2017; Gwizdka+ 2017);
Eye Movement Scanpath 1 Eye Movement Scanpath 2
Similar? Different?
How Much?
6. Background: Convolutional Neural Networks
Introduction User Study Scanpath Image Classification Interpretability Conclusion
• image classification is a major application of CNNs
• take an input image and predict a label for the image (e.g. “cat” or “dog”?)
• transfer learning: training received by a CNN for solving one task can be re-used to solve another related task
• e.g. training from cat/dog classifier can be re-used to classify traffic symbols
• benchmark CNN models, pre-trained on millions of images for classification tasks (ImageNet challenge) are readily available
• e.g. VGG, ResNet, DenseNet, etc.
Image: https://towardsdatascience.com/covolutional-neural-network-cb0883dd6529
7. Proposed Approach
Introduction User Study Scanpath Image Classification Interpretability Conclusion
Image: https://dev.to/frosnerd/handwritten-digit-recognition-using-convolutional-neural-networks-11g0
Scanpath - Image CNN Image Classifier
User
Perceived
Relevant?
Prediction
Eye movement
Scanpath
9. Experimental Design
Introduction User Study Scanpath Image Classification Interpretability Conclusion
• Participants (N = 25, college-age students)
Example Trigger Q: The submarine Kursk was part of which Russian fleet?
Perceived Relevant Perceived Irrelevant
Trigger
Question
TREC 2005
Q&A Task
Spacebar
Relevance
Judgement
(binary)
Y/N then
Spacebar
+
1s
Short News
Article
AQUAINT Corpus
of English News
Text
+
Fixation
>= 2s
13. Generating Scanpath-Images: Fixation Start Time
Introduction User Study Scanpath Image Classification Interpretability Conclusion
First
Saccade
Last
Saccade
Matplotlib’s winter colourmap
• each linearized saccade has a solid colour
Saccade Colour
16. Scanpath-Image Classification
Introduction User Study Scanpath Image Classification Interpretability Conclusion
Given only the scanpath-image of a user’s eye movements on
the news article, predict if the user perceived the article to be
relevant for answering the trigger question.
Image: https://dev.to/frosnerd/handwritten-digit-recognition-using-convolutional-neural-networks-11g0
Scanpath - Image CNN Image Classifier
Perceived
Relevance
Prediction
18. Scanpath-Image Classification: Results
Introduction User Study Scanpath Image Classification Interpretability Conclusion
For this specific task:
• Models do not overfit
• Shallow models classify better than deep models
Shallow
Deep
Table 1 from paper
20. Attempt to Interpret CNN Predictions
Introduction User Study Scanpath Image Classification Interpretability Conclusion
Gradient-Weighted Class Activation Mapping (Grad-CAM)
Original Image CAM for “Cat” classCAM for “Dog” class
2017 IEEE International Conference on Computer Vision
21. Attempt to Interpret CNN Predictions
Introduction User Study Scanpath Image Classification Interpretability Conclusion
SCANPATH CLASS ACTIVATION MAP (CAM) AVERAGE CAM
Across all scanpath-images in this relevance class
Perceived Irrelevant
Perceived Relevant
23. Conclusion
Limitations:
• very simple information search task
• short texts of similar type
• relatively uniform group of participants (college-age
students)
Future Directions:
• complex scenarios, e.g., freely searching on the open
web
• diverse participants, e.g., young vs. older adults
• Eye-movement scanpath-image
classification:
• no aggregate measures: all eye-tracking data is
used
• spatio-temporal aspects of eye-movements are
preserved
• knowledge of screen content not needed
• additional insights (e.g. reading / scanning) not
needed
• Proof of concept:
• promising results, even with small dataset, without
overfitting
• CNNs trained for a different task can detect
patterns in eye-movements which are concordant
with prior literature
Introduction User Study Scanpath Image Classification Interpretability Conclusion
24. Acknowledgements
Student Travel Grant
Experimental Design Contribution,
Data Collection
Prof. Bradley Hatfield
Dr. Rodolphe Gentili
Dr. Joe Dien
Hyuk Oh
Kyle James Jaquess
Li-Chuan Lo
Department of Kinesiology,
University of Maryland, College Park
For inspiration:
Blog post on using mouse
trajectories for fraud detection
Gleb Esman
Splunk Inc.
THANK
YOU
@NilavraBnilavra@ieee.org ixlab.ischool.utexas.edu
Full paper:
https://dl.acm.org/doi/10.1145/3343413.3377960
https://arxiv.org/abs/2001.05152