Learning Personalised Prices in Ad Auctions with Game Theory and Deep Learning

Spotify Research developed a method combining game theory and deep neural networks to optimize personalized reservation prices in ad auctions. The approach uses Mixture Density Networks with embedded Nash equilibrium constraints to infer advertisers' hidden willingness to pay from bidding data. Tested on 100,000 real auctions across ten markets, the system achieved an average 4% revenue increase (up to 11.8% in some markets) by setting dynamic prices based on user features and advertiser characteristics. The method bridges economic theory with machine learning to create more efficient ad marketplaces.