LO-FAR: New CPU method ranks ad features 100x faster without GPU retraining

Industrial ad recommendation models often rely on hundreds of sparse, high-cardinality ID-list features (user histories, context IDs), each with its own embedding table. These features dominate storage and training costs, so choosing which ones to keep is a recurring systems-level decision constrained by compute budgets and iteration speed. Traditional ranking methods – permutation importance, stochastic gates (BinaryST), or Shapley values – require GPU-bound retraining loops that can take days.

Enter LO-FAR (Localized Feature Ranking): a CPU-only, model-agnostic workflow that evaluates each candidate feature by its standalone hold-out predictive signal using lightweight local estimators. On a dataset of over 1 million logged interactions with 475 features, LO-FAR finishes ranking in roughly 2 CPU-hours. Its downstream Normalized Entropy gains on CTR and CVR tasks are competitive with shuffle-based importance, BinaryST, and a coverage heuristic across budgets of 100 to 400 retained features. The paper (RecSys 2026) demonstrates that when cost and turnaround time are binding constraints, a simple local filter can outperform heavier interaction-aware methods in production.