Speeding Up Inference on Graph Models for Warm Start (AIRI)

Evgeny Frolov24 May 2025

Problem

A recommender is retrained, say, once a day, but users keep clicking between two trainings. A new user, or a known user with fresh clicks, has no up-to-date embedding, so the system uses a random vector or waits for the next training. Transformers avoid this: they take the new click sequence as input. Matrix factorization (MF) and graph neural networks (GNN) tie each user to an ID-specific embedding, and retraining a GNN is far too slow for this. Hybrid models are as slow as their graph part.

Example from the article: on MovieLens, full retraining of the graph part for test users takes about 4 orders of magnitude longer than a fold-in update in the hybrid model.

Solution in brief

Fold-in: freeze all item and other user embeddings and re-solve the optimization for one user's embedding. For graph models the problem is of fourth order with no closed form, and it is solved with ADMM or quasi-Newton methods, or analytically through UltraGCN. Quality stays close to full retraining at a small overhead over doing nothing.

Questions this case answers
  • How to update a graph-based recommender with a new user's interactions between full retrainings without retraining the model
  • How to keep the graph component from slowing down a hybrid recommender at serving time
  • How to update a graph-based recommender for a user's new interactions between retrainings without retraining the model
  • How to build recommendations for a new or returning user in a matrix factorization or GNN model in real time

Tools and methods

  • Fold-in — update one user's embedding against frozen item embeddings
  • UltraGCN — graph model with an analytical fold-in
  • LightGCN, NGCF — standard graph collaborative-filtering models
  • ADMM, quasi-Newton methods — explicit solvers
  • SASRec plus a graph part — transformer and graph hybrid
Continue reading — it's free

Google or a one-time email code. No password.