Query Rewards

• Pinterest uses PinnerSAGE to generate clusters of user's engaged pins
• PinnerSAGE clusters are sampled as the source of queries for selection
• No feedback is used for future query selection
• Query Reward is a new component added for feedback loop
• Query Reward computes engagement rate of each query for future selection
• Future reward gradually drops for clusters with no engagement
• Query Reward allows for a more responsive and instant feedback loop

User Profiling with PinnerSAGE Overview

• PinnerSAGE generates clusters of user's engaged pins
• Clusters are based on pin embedding by grouping nearby pins together
• Sampling is done from PinnerSAGE clusters as source for queries
• Downstream engagements from last request's sampling results are not considered

Homefeed Query Composition After Adding Query Reward

• Query Reward workflow computes engagement rate of each query
• Leads to better selection of queries for candidate generation
• Allows for feedback loop based on users' engagement rates
• Cluster weight given an average when no engagement is present