From Clicks to Conversions: Architecting Shopping Conversion Candidate Generation at Pinterest
Here's a 3-sentence summary of the blog post on Archtecting Shopping Conversion Candidate Generation at Pinterest:
Pinterest engineers developed a dedicated candidate generation model for shopping conversions, improving advertiser performance and click-through rates by 11%. To address the sparsity of conversion data, the team employed a multi-surface model, dual positive signals, and negative sampling, as well as a multi-task approach with engagement prediction as an auxiliary task. The feature engineering included a two-tiered approach to capture user-side and Pin-side features, while the model architecture used a two-tower design with DCN v2 and parallelized cross layers to achieve higher recall and online metric gains.
Machine LearningData