When Recommendations Systems Go Bad
summary
Covers concrete failures: a hotel recommender that split Mac vs PC users after finding Mac sessions more likely to book 4–5 star hotels (three alternative hotels shown on the hotel page), Amazon-style co-purchase noise that recommends illicit items after a kitchen scale purchase, Google ads showing “criminal record” links for Black-sounding names, and Google Photos tagging a Black couple as gorillas. Presents fixes actually used: logistic regression for interpretability, data segregation via ensemble modelling, and adFisher generated fake profiles as test datasets to detect unintended bias.
AI-written summary · about
chapters (11)
- 0:00Start
- 1:40National Day of Civic Hackin
- 1:53Tech Meetups on the Rise
- 3:02Recommendation Systems: Collaborative F
- 3:43Recommendation Systems: Rating Predicti
- 4:15Recommendation Systems: Learning To Ra
- 13:43Women Who Code NYC
- 15:06Logistic Regression
- 15:42Feature Engineering and Interactions
- 16:42Ensemble Model, Data Segregation
- 18:07Diversity Controlled Testing
related talks
41mRecommender Systems: Crafting Effective Personalised Experiences
23mFrom Data to Recommendations: Building an Intelligent System with ROR
32mLet's talk about Recommendation Systems!
15mFrom Personalized to Programmed
28mBreaking the Chains of Oppressive Software
41mGhost in the Machine: The Unexpected Results of Not Understanding Your Users
43mKeynote: Consequences of an Insightful Algorithm
47mAlgorithmic bias is everywhere (especially at Breeze) - what can we do about it?