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

TL;DR

This paper proposes a method that dynamically updates customer profiles and composes a feed every two minutes, employing precomputed embeddings and their respective similarities, and enhanced customer engagement and experience.

Abstract

In personalized recommender systems, embeddings are often used to encode customer actions and items, and retrieval is then performed in the embedding space using approximate nearest neighbor search. However, this approach can lead to two challenges: 1) user embeddings can restrict the diversity of interests captured and 2) the need to keep them up-to-date requires an expensive, real-time infrastructure. In this paper, we propose a method that overcomes these challenges in a practical, industrial setting. The method dynamically updates customer profiles and composes a feed every two minutes, employing precomputed embeddings and their respective similarities. We tested and deployed this method to personalise promotional items at Bol, one of the largest e-commerce platforms of the Netherlands and Belgium. The method enhanced customer engagement and experience, leading to a significant 4.9% uplift in conversions.

Figures (5)

• Figure 1: User-to-item framework: Single vectors from the user encoder limit representation and interpretability. Keeping them fresh demands high-maintenance infrastructure.
• Figure 2: Query-to-item framework: Query embeddings and their similarities are precomputed. Users are represented by a dynamic set of queries that can be updated as needed.
• Figure 3: The major steps involved in generating near real-time personalized recommendations
• Figure 4: The SIMO (Single Input Multi Ouput) embedding model generates three embeddings per item in one model run using three special tokens: [Q_V], [Q_B], and [TGT].
• Figure 5: Inputs to dual SIMO encoders: the query encoder takes in the metadata of the query item and generates three embeddings and the target encoder takes in the metadata of the target item and generates three embeddings. During training, the loss is determined by the target embedding derived from the target item $t_i$ encoder and pairing it with a query embedding from the query item $q_i$ encoder, selected by the relation $r_i$ indicator.