DBTB INT Oscar Celma r
summary
How Pandora balances playing familiar favorites with serendipitous discovery using real user signals and engineering checks is described concretely here including 65 billion thumbs of feedback, a user base of 18 million active users, and a roughly 20 person data science team. The workflow uses Python and Spark, produces offline metrics like AUC, applies offline quality control checks to catch obviously wrong recommendations, then runs A/B tests to measure real user response and iterate until models both score well and sound sensible.
AI-written summary · about
related talks
35mBuilding Data-Intensive Teams
20mDataEngBytes 2023 - PER-T1-06 - Kamal Hossain
45mData as culture as software
49mdbt: A Software Engineering Approach To The Analytics Workflow
37mData Poker not Data Chess - thinking in Bets for Data Teams
16mExploring Customer Data Platforms and Testing
13mData for product
45mChris Harbert: Testing AI at the Scale of Big Data