A naukri.com initiative
Medium
1M
260
Image Credit: Medium
Next-Level Personalization: How 16k+ Lifelong User Actions Supercharge Pinterest’s Recommendations
- Pinterest's home feed personalization process involves retrieving candidate pins based on user interests and ranking them using the Pinnability model.
- TransActV2, an upgraded model, addresses the limitation of not being able to model a user's lifelong behavior on Pinterest.
- TransActV2 leverages long user sequences, a Next Action Loss function, and efficient deployment solutions.
- It aims to capture evolving user interests, provide rich personalization, and handle the challenges of processing extensive user histories.
- The model can now handle up to 16,000 user actions and includes features like timestamp, action type, action surface, and PinSage embeddings.
- Nearest Neighbor Selection is used to reduce the length of user sequences during ranking, improving efficiency.
- The model architecture involves representation layers, a Transformer Encoder, and downstream heads for various action predictions.
- Next Action Loss (NAL) is introduced to enhance action forecasting by predicting the user's next action given the context and history.
- Efficient serving and deployment strategies like Nearest Neighbor feature logging and custom Triton kernels are implemented to handle lifelong sequences.
- TransActV2's improvements led to significant enhancements in offline metrics and user engagement, outperforming prior systems.
- Real-world A/B tests showed substantial increases in repins, reduced hide signals, improved session quality, and enhanced content diversity.
Read Full Article
15 Likes
For uninterrupted reading, download the app