Probe-Enhanced Drift-Aware Model Updating for Compute-Efficient Sequential Recommendation

Sequential recommenders must adapt to new interactions without repeatedly paying the cost of full-history retraining. Strict drift-triggered updating appears attractive, but can inadvertently suppress useful learning when marginal item frequencies are stable. We develop DAGR-P, a probe-enhanced update scheduler that performs a one-epoch new-data probe when a permutation drift screen is negative and allocates a longer, drift-dependent fine-tuning or recency-replay update otherwise. Candidate deployment follows a paired per-user validation rule. We implement the recency-replay update and the paired admission gate of the SPROUT compute-aware framework, and examine an additional user-matched drift detector. Five independent random seeds are evaluated on each of three explicitly synthetic, globally ordered interaction processes (22,680 interactions, 420 users, and 144 items per seed). Under stationary and abrupt processes, DAGR-P improves over a strict-gate policy from 0.1756 to 0.1857 and from 0.2540 to 0.2588 NDCG@10, respectively. Under gradual drift it obtains 0.2773, compared with 0.2689 for full retraining, with 32.93M versus 118.39M normalized training-score equivalents. A strong new-only updater remains more accurate in all three processes. The study identifies when low-budget probes counteract missed updates, and also shows why matched-user drift tests can be overly conservative. These results support a transparent compute–accuracy analysis, not a claim of field-scale effectiveness.