Improving Quality of Recommended Content through Pinner Surveys

Pinterest describes how Pinner (user) surveys are used to train a machine learning model that improves content quality recommendations across Homefeed, Related Pins, and Search.

• •5,000 Pins were rated by Pinners on a 1–5 visual appeal scale, with at least 10 responses per image to reduce noise, sampled across five L1 interest verticals
• •Survey data was used to train a 92k-parameter fully-connected neural network using pairwise ranking, predicting which of two images is visually higher quality within the same interest vertical
• •The model outputs a per-Pin quality score that is incorporated as a feature in ranking systems to boost high-quality content and suppress low-quality or clickbait content
• •Results showed positive outcomes for both user wellbeing metrics and business engagement, demonstrating a win/win alignment between user satisfaction and platform goals