History compression stops being recomputed per candidate: X-split ARGUS lifts conversion value 4.65% in a live bidding system
Synopsis
Click-through rate prediction needs long behavioral histories, yet re-encoding that history on every ranking request strains production compute and latency budgets. X-split ARGUS compresses history into a set of tokens that do not depend on the candidate, computes them offline and reuses them across requests, while the online ranker cross-attends to those tokens and to recent events for each candidate. Offline normalized entropy falls 0.81% relative to Context-aware ARGUS, retaining more than 70% of the uncompressed teacher's improvement; in an online A/B test conversion value rises 4.65% and clicks 2.96%, at the cost of 15% more online GPUs of the same type for the same traffic.
Interpretation
Long behavioral history is no longer encoded from scratch on every ranking request; it is compressed into a set of history tokens reusable across candidates, and the online path handles only recent events and the candidate ad. The prior Context-aware ARGUS is a two-tower ranker whose user tower holds both a history compressor and a runtime encoder, with the candidate ad represented by the item tower and scored by a dot product. On internal real-time bidding logs covering 365 days of history, X-split ARGUS reaches a normalized entropy of 0.7146 against 0.7204 for Context-aware ARGUS, a 0.81% relative reduction; the uncompressed teacher reaches 0.7123, a 1.136% relative reduction, so the compressed student retains 71.6% of the teacher's improvement.
The history compressor and the online ranker are trained end to end, and because the compressor's tokens do not depend on the candidate they can be precomputed offline and reused across requests. The compressor is a Transformer with causal self-attention that makes one pass over historical events and keeps the last contextualized states of a prefix as the compressed representation, while a runtime encoder processes recent events with a sliding window. The ranker uses the candidate as queries and the compressed tokens plus recent-event states as keys and values in cross-attention; during training a larger teacher that processes uncompressed history supplies per-candidate click logits through a distillation head with weight 0.15, and the teacher stays out of the online scoring path.
Perspective
Results come from the largest advertising surface of this real-time bidding system, about 50% of its total traffic, with a seven-day A/B test assigning 6% of that surface's traffic to each of treatment and control; the offline comparison uses an internal industrial dataset covering 365 days of history, with a one-day train-test gap and a three-day test period.
The fixed checkpoint's relative advantage declines approximately linearly with the gap between training data and prediction, about 0.05 percentage points per day, so the deployed model is fine-tuned daily on targets from the most recent three days and still leaves roughly one to two days between the latest training events and serving; in the depth ablation, deeper encoders change capacity and context propagation together, and the two are not separated.
