3. INTRODUCTION
▸ Flipboard - Content Recommendation
▸ Background in Computer Science and
Machine learning.
▸ Meetup Co-organiser - AI in
Production and PyData.
▸ Interests: Recommender Systems,
Natural Language Processing,
Counterfactual Reasoning and
Bayesian Machine Learning
▸ “Real” interests: Sarcasm and Football
— Liverpool FC and Vancouver
Whitecaps.
4. ECHO CHAMBERS
ECHO CHAMBERS / FILTER BUBBLES
▸ An environment where readers are
only shown content that reinforces
their views without exposing them
to different opinions.
▸ Not a new phenomenon.
▸ Amplified through personalization
and recommender systems.
6. ECHO CHAMBERS
WHY PERSONALIZATION
▸ Personalized recommendations
allow users to find information that
is aligned with their interests
without needing them to sift
through pages of irrelevant content.
▸ Personalization allow companies to
capture users’ “true” interests and
align them with more relevant
advertising.
7. ECHO CHAMBERS
DRAWBACKS OF PERSONALIZED RECSYS
▸ Algorithmic gatekeeping: Filtering
out potentially important content.
▸ Polarised views: People expressing
genuine disbelief that fellow
citizens could possibly
countenance alternative opinions.
▸ Lack of access to historical
recommendations for analysis and
research. (Carole Cadwalladr - TED
Talk 2019)
8. A SQUIRREL DYING IN YOUR FRONT
YARD MAY BE MORE RELEVANT TO
YOUR INTERESTS RIGHT NOW THAN
PEOPLE DYING IN AFRICA.
Mark Zuckerberg
ECHO CHAMBERS
9. RECOMMENDER SYSTEMS
▸ An attempt to narrow down
content for consumption on the
internet.
▸ Nearly everything on the internet
uses one — Google Search,
Twitter, Facebook, Amazon
Product Recommendations,
Netflix, Quora, Spotify etc.
▸ Personalisation helps with user
engagement.
10. RECOMMENDER SYSTEMS
DIFFERENT TYPES OF RECOMMENDER SYSTEMS
▸ Content-based Filtering
▸ Tag each piece of content with some features (topics / genres) .
▸ Collect users’ interests implicitly or explicitly (or both).
▸ Match content with users based by some definition of similarity.
▸ Collaborative Filtering
▸ Purely depend on usage, and understand patterns of similar users on
similar items.
▸ Exploration / Exploitation
11. RECOMMENDER SYSTEMS
CONTENT-BASED FILTERING
▸ Content-based filtering can generally give users
more control in terms of what they want to see.
▸ Explicit Features can be collected.
▸ Domains, Authors, Producers, Singers etc.
▸ Inferred Features can be derived.
▸ Keyword/topic extraction, content
classification etc.
▸ A user’s affinity to these features in
accumulated based on historical usage.
▸ Content with features that are the most similar
to a user’s derived affinity is recommended.
12. RECOMMENDER SYSTEMS
COLLABORATIVE FILTERING - SOCIAL / CROWDSOURCED RECOMMENDATIONS
▸ Derive patterns purely
from usage.
▸ User-user similarity:
Recommend content that
similar users “liked”.
▸ Item-item similarity:
Recommend a similar to
item to the one that the
user just “liked”.
13. RECOMMENDER SYSTEMS
PROS / CONS (PURELY FROM A RECSYS PERSPECTIVE)
- Content based filtering:
- Requires developing and maintaining feature extractors — errors can
propagate.
- It can however be decoupled from usage.
- Collaborative filtering:
- Reliant on usage, less technical debt and generally provides more engaging
recommendation for items with a lot of usage.
- It doesn’t work for new content, can be click-baity.
- More opaque to the end-users.
▸ Most common fix - Use a hybrid method that does both content based filtering and
collaborative filtering.
14. RECOMMENDER SYSTEMS
RECSYS AS MATRIX FACTORIZATION - CBF
Word Counts
Content
Documents Documents
Features
Content
Features
Lee, D. D., & Seung, H. S. (2001). Algorithms for non-negative matrix factorization. In Advances in neural information
processing systems (pp. 556-562).
15. RECOMMENDER SYSTEMS
RECSYS AS MATRIX FACTORIZATION - CF
Document - Usage
Users
Documents Documents
Factors
Users
Factors
Hu Y, Koren Y, Volinsky C. Collaborative filtering for implicit feedback datasets. In Data Mining, 2008. ICDM'08. Eighth IEEE
International Conference on 2008 Dec 15 (pp. 263-272). Ieee.
16. ECHO CHAMBERS IN RECOMMENDER SYSTEMS
SO WHY DO ECHO-CHAMBERS FORM?
▸ In CBF - Users are only exposed to
what the system thinks is
“relevant” to the user. Most user-
interfaces do not encourage
exploration.
▸ In CF - Users are only exposed to
either items that are already similar
to what they consume, or items
that users similar to them read.
▸ Diversity, importance, quality and
many similar features are often
ignored.
Williams, H. T., McMurray, J. R., Kurz, T., & Lambert, F. H.
(2015). Network analysis reveals open forums and echo
chambers in social media discussions of climate
change. Global Environmental Change, 32, 126-138.
18. ECHO CHAMBERS IN RECOMMENDER SYSTEMS
▸ This is not purely a software engineering or a machine
learning problem.
▸ But…
19. ECHO CHAMBERS IN RECOMMENDER SYSTEMS
EXPLORATION?
▸ How about instead of always
exploiting what we know about
the user, we always show some
random content to the user? —
Bubble Popping
▸ Unfortunately, this requires
user education or design
support otherwise it can
confuse an end-user.
20. ECHO CHAMBERS IN RECOMMENDER SYSTEMS
A DATA-SCIENCE THOUGHT
▸ Machine learning algorithms generally try to optimize for some metric.
▸ Some recent work at ICLR/ICML/NeurIPS on treating this as a
constrained optimization problem.
http://www.progressfocused.com/2016/11/coming-out-of-our-echo-chambers.html
Donini, M., Oneto, L., Ben-David, S., Shawe-Taylor, J. S., & Pontil, M. (2018). Empirical risk minimization under fairness
constraints. In Advances in Neural Information Processing Systems (pp. 2791-2801).
21. ECHO CHAMBERS IN RECOMMENDER SYSTEMS
ROUNDUPS
▸ A single story from one-source
doesn’t do important stories justice
— Facebook Trending stories being
a prime example.
▸ Flipboard, Google News,
TechMeme and many others now
try to show different publications’
take on the same story.
22. ECHO CHAMBERS IN RECOMMENDER SYSTEMS
ROUNDUPS - II
https://
engineering.flipboard
.com/2017/02/
storyclustering
23. ECHO CHAMBERS IN RECOMMENDER SYSTEMS
ROUNDUPS - III
▸ This has been live on Flipboard
since Feb 2017.
▸ Allows people to explore if they
want to.
▸ Worst case - People ignore the
other perspectives.
24. ECHO CHAMBERS IN RECOMMENDER SYSTEMS
HUMAN IN THE LOOP
▸ Use CBF and CF outputs as features along with editorial heuristics
and journalistic features in a LearnToRank system.
25. ECHO CHAMBERS IN RECOMMENDER SYSTEMS
SOCIAL AND TRUST BASED RECOMMENDER SYSTEMS
▸ Given two individuals - the source (Node A) and sink
(Node C), derive how much trust is between them
given some social connections.
▸ Discover bipartisans/gatekeepers.
▸ Algorithms:
▸ Advogato (Levien et. al.), Appleseed (Ziegler et. al.),
MoleTrust (Massa et. al.), TidalTrust (Goldbeck et. al.)
A B C
Garimella, K., De Francisci Morales, G., Gionis, A., & Mathioudakis, M. (2018, April). Political discourse on social media: Echo
chambers, gatekeepers, and the price of bipartisanship. In Proceedings of the 2018 World Wide Web Conference on World Wide
Web (pp. 913-922). International World Wide Web Conferences Steering Committee.
26. ECHO CHAMBERS IN RECOMMENDER SYSTEMS
AND MANY MORE…
▸ Diversity, fairness, balance, explainability and many other
issues are often raised in academia.
▸ See References for more discussion.
Chaney, A. J., Stewart, B. M., & Engelhardt, B. E. (2017). How algorithmic confounding in recommendation systems increases
homogeneity and decreases utility. arXiv preprint arXiv:1710.11214.
Sanz-Cruzado, J., & Castells, P. (2018, September). Enhancing structural diversity in social networks by recommending weak ties.
In Proceedings of the 12th ACM Conference on Recommender Systems (pp. 233-241). ACM.
Bias and Fairness in AI, AI for Social Good, NeurIPS, Montreal, 2018. https://www.youtube.com/watch?v=ljBNmWYC7rg
27. CONCLUSION
CONCLUSION
▸ Algorithms are the gatekeepers of information on the
internet. We need to encode journalistic ethics into our
recommender systems.
▸ Recommender Systems are quite simple, and it should not
be a black box.
▸ Algorithms should be held accountable but there are
things that companies can do to mitigate some of these
issues.
28. ECHO CHAMBERS IN RECOMMENDER SYSTEMS
REFERENCES
▸ Eli Pariser: “The Filter Bubble”, TED Talk: https://www.ted.com/talks/eli_pariser_beware_online_filter_bubbles/up-
next?language=en
▸ Carole Cadwalladr TED Talk: https://www.ted.com/talks/
carole_cadwalladr_facebook_s_role_in_brexit_and_the_threat_to_democracy/up-next?
utm_source=twitter.com&utm_medium=social&utm_campaign=tedspread
▸ Garimella, K., De Francisci Morales, G., Gionis, A., & Mathioudakis, M. (2018, April). Political discourse on social
media: Echo chambers, gatekeepers, and the price of bipartisanship. In Proceedings of the 2018 World Wide Web
Conference on World Wide Web (pp. 913-922). International World Wide Web Conferences Steering Committee.
▸ Chaney, A. J., Stewart, B. M., & Engelhardt, B. E. (2017). How algorithmic confounding in recommendation systems
increases homogeneity and decreases utility. arXiv preprint arXiv:1710.11214.
▸ NeurIPS 2017 Workshop: https://www.k4all.org/event/prioritising/
▸ What Algorithms can learn from Journalism: https://about.flipboard.com/inside-flipboard/what-algorithms-can-
learn-from-journalism/
▸ Xavier Amatriain, RecSys 2017 Summer School: https://www.slideshare.net/xamat/recommender-systems-in-
industry