# Pfeed: Generating near real-time personalized feeds using precomputed embedding similarities

**Authors:** Binyam Gebre, Karoliina Ranta, Stef van den Elzen, Ernst Kuiper, Thijs Baars, Tom Heskes  
**arXiv:** [2402.16073](https://arxiv.org/abs/2402.16073)  
**Date:** 2024-02-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.16073v2/

## Key points

- Personalized recommender systems commonly use embeddings to encode customer actions and items
- Retrieval is done in the embedding space through approximate nearest neighbor search
- Challenges include restrictions on diversity within user embeddings and costly real-time infrastructure maintenance
- A method was proposed to address these challenges by dynamically updating customer profiles and generating personalized feeds every two minutes using precomputed embeddings
- The method was implemented at Bol, an e-commerce platform, resulting in enhanced customer engagement and a 4.9% uplift in conversions
- Historical customer interactions data is used with an encoder model to generate query embeddings for items, which are matched against item embeddings using ANNS for recommendations
- The system incorporates ranking algorithms, diversity considerations, business rules integration, and merging recommendations based on recent interactions
- By updating near real-time every two minutes, the method successfully delivers tailored feeds while addressing concerns about diverse user interests and minimizing infrastructure costs

## Summary

In personalized recommender systems, embeddings are commonly utilized to encode customer actions and items. Retrieval is then conducted in the embedding space through approximate nearest neighbor search. However, this approach presents challenges such as potential restrictions on the diversity of captured interests within user embeddings and the necessity for costly real-time infrastructure to maintain their up-to-date status. To address these issues in a practical industrial setting, a method was proposed that dynamically updates customer profiles and generates personalized feeds every two minutes by leveraging precomputed embeddings and their corresponding similarities. This method was implemented and tested at Bol, a prominent e-commerce platform serving the Netherlands and Belgium, to personalize promotional items. The results were promising with enhanced customer engagement and experience leading to a notable 4.9% uplift in conversions. The process involved utilizing historical customer interactions data alongside an encoder model to generate query embeddings for items in the catalogue. These query embeddings were then matched against item embeddings stored in an index vector database using approximate nearest neighbor search (ANNS) to retrieve relevant recommendations. Furthermore, the system incorporated elements of ranking algorithms, diversity considerations, business rules integration, and merging recommendations based on recent customer interactions. By combining these components effectively within a near real-time framework that updates every two minutes, the method successfully delivered tailored feeds to users while addressing concerns related to maintaining diverse user interests and minimizing infrastructure costs associated with real-time updates. Authors Binyam Gebre, Karoliina Ranta, Stef van den Elzen, Ernst Kuiper, Thijs Baars, and Tom Heskes present this innovative approach in their paper titled "Pfeed: Generating near real-time personalized feeds using precomputed embedding similarities," offering valuable insights for enhancing personalization strategies in e-commerce platforms through efficient use of embedding techniques and dynamic feed generation processes.