How Pinterest Uses AI to Supercharge Its Recommendations System
Summary
Pinterest, a social network known for its visual discovery and planning website, has developed an advanced recommendation system powered by artificial intelligence (AI). This system, known as PinSage, uses a graph convolutional network (GCN) to learn about pins or nodes in web-scale graphs, enhancing the platform’s ability to provide relevant recommendations to its users. By leveraging AI, Pinterest aims to continuously improve its recommendation capabilities, keeping pace with its growing user base and vast collection of saved objects.
The Challenge
With over 200 million monthly active users and more than 100 billion saved objects, Pinterest faces the challenge of providing accurate and relevant recommendations to its users. Traditional deep learning models are limited in their ability to handle such large-scale data, necessitating the development of more advanced frameworks like PinSage.
The Solution: PinSage
PinSage is a deep learning-based framework that uses a graph convolutional network (GCN) to learn about pins or nodes in web-scale graphs. Each pin is associated with an image and a set of textual annotations such as the title and description. This framework allows Pinterest to analyze more information than standard deep learning models, providing context and disambiguating pins that are visually similar but semantically different.
Training PinSage
To train PinSage, the team used NVIDIA Tesla GPUs and the cuDNN-accelerated TensorFlow deep learning framework. The graph convolutional network was trained on approximately 18 terabytes of data, or 7.5 billion examples on a graph, with 3 billion nodes representing pins and boards.
Key Applications
PinSage has several key applications within Pinterest’s services, including:
- Recommendation of Pins: Including dynamic pins like those for ads and shopping.
- Classification and Clustering: Essential for tasks like Related Pins, Search, Shopping, and Ads.
- Reranking: Improving the relevance of recommendations based on user interactions.
Impact
The development of PinSage marks the largest application of deep graph embeddings to date, paving the way for a new generation of web-scale recommender systems based on graph convolutional architectures. This framework has the potential to supercharge the backend of product discovery and power the future of shopping.
Technical Overview
The technical overview of PinSage highlights its ability to learn from vast amounts of data, leveraging graph information to provide context and disambiguate similar pins. This capability is crucial for improving the accuracy and relevance of recommendations, enhancing user engagement and overall experience.
Table: Key Features of PinSage
| Feature | Description |
|---|---|
| Graph Convolutional Network (GCN) | Learns about pins or nodes in web-scale graphs. |
| Data Training | Trained on 18 terabytes of data, or 7.5 billion examples. |
| Applications | Recommendation of Pins, Classification, Clustering, Reranking. |
| Impact | Largest application of deep graph embeddings, enhancing product discovery and shopping. |
Further Insights
The development of PinSage underscores the importance of AI in modern recommendation systems. By leveraging advanced deep learning frameworks, platforms like Pinterest can provide more accurate and relevant content to their users, driving engagement and growth. This approach sets a precedent for future innovations in web-scale recommender systems.
Conclusion
Pinterest’s use of AI in its recommendations system, specifically through the development of PinSage, demonstrates the power of advanced deep learning frameworks in enhancing user experience. By continuously improving its recommendation capabilities, Pinterest aims to provide more accurate and relevant content to its users, driving engagement and growth.