Claudia Perlich, Chief Scientist, Dstillery, NYC 2017
summary
Models chase the strongest signal and can mislead when data mixes different behaviors. Using logistic regression on an anonymized Facebook-style likes/age set produces extreme ranks: top 1% → 75% men with basic features and 100% men after adding likes. Bot traffic and accidental events distort outcomes — bot-generated visits grow from ~6% to ~30% and accidental ad clicks are about 40%, so top-ranked targets end up dominated by bots or accidents rather than genuine intent. Models stay confident even with heavy label noise.
AI-written summary · about
chapters (11)
- 0:00Start
- 3:10Take 1: Predict Gender Based on age ...
- 3:25Gender based on age: very little signal ...
- 10:30Witness a spike in human predictability ..
- 11:31Percent bot traffic on conversion metrics
- 17:14Potentially three populations in the location prediction
- 18:20Real Estate
- 19:00How much randomness can a model absorb?
- 21:01Where do we find frequent traveler?
- 22:04Predict who goes to JFK?
- 23:39Beware of (unintentional) discrimination based on predictability
related talks
13mClaudia Perlich, Chief Scientist, Dstillery
30mPredictive Analytics in Social Media and Online Display Advertising
15mUnderstanding Decisions Driven by Big Data
22mCommunicating Algorithmic Process in Online Behavioral Advertising
43mKeynote: Consequences of an Insightful Algorithm
32mPlatform-supported Auditing Of Social Media Algorithms For Public Interest
22mGenerating a Billion Personal News Feeds
54mOn How Machine Learning and Auction Theory Power Facebook Advertising