A naukri.com initiative
Medium
4w
222
Image Credit: Medium
Unlocking Efficient Ad Retrieval: Offline Approximate Nearest Neighbors in Pinterest Ads
- In the realm of advertising, Pinterest explores the use of Offline Approximate Nearest Neighbors (ANN) for efficient ad retrieval alongside Online ANN.
- Offline ANN is beneficial for high throughput, low-latency query responses, and static query contexts in large-scale data operations, as discussed in the article.
- Challenges faced by Pinterest with expanding ads inventory led to transitioning to the Inverted File algorithm for improved ANN efficiency.
- The architecture of Online/Offline ANN retrieval involves real-time online serving systems and batch offline query embedding storage.
- Advantages of Offline ANN include cost efficiency and extensibility, while limitations include real-time processing constraints and fixed neighbor numbers.
- Offline ANN is suitable for stable query contexts with reduced cost priorities, while Online ANN is preferred for real-time processing needs.
- Pinterest's application of offline ANN includes use cases like similar item ads and visual embedding for improved ad relevance and lower infrastructure costs.
- For similar item ads, offline ANN showed lower infra costs and better engagement compared to online ANN.
- Visual embedding with offline ANN had comparable candidate fetch rates and better performance metrics at reduced infra costs.
- Future plans include integrating offline ANN into other interfaces and developing a Pinterest-specific offline ANN framework for enhanced scalability and features.
Read Full Article
13 Likes
For uninterrupted reading, download the app