FORTE: A Production Framework for Feature Evaluation and Deprecation in Industrial Recommender Systems
FORTE removed nearly half the features from one recommender model without an online regression.
The paper describes a production framework for finding low-value recommender features, testing them in grouped ablations, and tracing downstream model use before removal. In a search-relevance model, FORTE produced candidate sets of 32, 48, and 59 features; the largest failed online guardrails, while the 48-feature set was chosen for deprecation. In a larger ranking model, eight removal candidates passed retraining and A/B testing, cutting trainable parameters by 29.9% and improving serving throughput by 19.8-25.5%. Source: ArXiv · AI/CL/LG's note.
The paper describes a production framework for finding low-value recommender features, testing them in grouped ablations, and tracing downstream model use before removal. In a search-relevance model, FORTE produced candidate sets of 32, 48, and 59 features; the largest failed online guardrails, while the 48-feature set was chosen for deprecation. In a larger ranking model, eight removal candidates passed retraining and A/B testing, cutting trainable parameters by 29.9% and improving serving throughput by 19.8-25.5%. Source: ArXiv · AI/CL/LG's note.
score 5