Uniboost — SIGIR '26 Poster

Abstract

We propose Uniboost, a unified traffic allocation framework for the recommendation blending stage. It introduces posterior value alignment to calibrate abstract model scores to anchor metrics with explicit business semantics, and employs independent linear boosting to decouple complex weighting schemes for precise attribution. Online A/B tests on a large-scale industrial platform show that Uniboost improves allocation efficiency at the micro level and provides macro-level guidance for system iteration.

1 Background & Challenges

The Blending stage merges candidates from diverse sources (ads, organic, cold-start) and allocates traffic via weighting mechanisms. Existing approaches suffer from three core problems:

Score Inflation — accumulated PID and Boost weights make scores explode, losing physical semantics and interpretability

Plan Coupling — multiple allocation plans (guaranteed delivery, cold-start, promotions) interfere with each other, making optimisation intractable

No Attribution — individual plan cost and contribution cannot be isolated, blocking data-driven decisions

Our Solutions

1

Posterior Value Alignment — identify Effective Completion Rate as anchor metric, turning abstract scores into business-semantic values without altering ranking order

2

Unified Boosting Paradigm — independent linear boosting with bias, covering existing schemes while keeping every weighted score interpretable

3

Precise Attribution — linear aggregation quantifies each plan's contribution, enabling ROI-based macro guidance

2 Uniboost Method

2.1 Value Alignment

Raw scores y′_v lack business semantics. We map them into anchor-metric space:

y_v = y′_v · μ_anchor ⁄ μ_score Eq. 1 — calibrates scores to expected Effective Completion Rate

2.2 Independent Boosting

s_p,v = 𝟙_p(v) · (w_p · y_v + b_p) Eq. 2 — each plan computed independently with learnable w_p, b_p

2.3 Linear Aggregation

r_v = y_v + ∑_p∈S s_p,v Eq. 3 — additive composition ensures full attributability

Backward Compatibility

w_ad=0, b_ad=PID → original guaranteed delivery is preserved.
w_cs≠0, b_cs=0 → conventional cold-start Boost. Unified with zero disruption.

Figure 1: Uniboost system overview — candidates flow through Model Scoring → Value Alignment → Independent Boost Calculation → Linear Aggregation → Exposure. A near-line/offline subsystem performs Distribution Tracking and ROI Analysis for closed-loop optimisation.

3 Experiments & Results

Large-scale online A/B test on Taobao Content Feeds. Control: original pipeline (align after weighting). Treatment: Uniboost with Effective Completion as anchor.

+1.69%

VV
(impressions)

+3.07%

Valued VV
(played > 3s)

+0.65%

Duration
(total watch time)

+2.54%

Valued Score
(composite UX)

RQ1 Micro-level Efficiency

Ad boosting score

−92.21%

Video boosting score

−95.81%

Weighting magnitudes drop by more than 90% while exposure shares stay stable — decoupled plans do not sacrifice delivery.

Insight: shrinking weighted-score magnitude removes unintended cross-plan interference, yielding measurably more efficient micro-level allocation.

RQ2 Macro-level Guidance

ROI-based offline analysis quantifies each plan's cost against its benefit:

ROI^VV_p = ΔVV_p ⁄ Cost_p

Plan-A was flagged as lowest ROI; the ablation confirms it:

Setting	VV	Valued VV	Duration	Valued Score
w/o Plan-A	+0.95%	+2.83%	+3.49%	+4.13%

Insight: per-plan cost/benefit dashboards turn traffic governance into a data-driven loop, accelerating strategy iteration.

4 RQ3 Anchor Metric Selection

A usable anchor must satisfy two principles. We screen six posterior metrics against the blending score (x axis: model score; each panel shares the same plot geometry):

Principle 1 Stability

Values stay uniformly distributed across every model-score interval — no score-dependent divergence or bias.

Principle 2 Low Calibration Error

Calibration error remains low and flat across all score intervals, so alignment introduces no systematic bias.

Eff. Completion y: 0 – 1.0 · stable, well spread

Duration y: 0 – 0.9 · long-tail divergence

Click y: 0 – 1.2×10⁻⁴ · long-tail divergence

Interaction y: 0 – 0.18 · sparse & volatile

All panels: x axis = blending score, y axis = posterior metric value (range noted per panel).

Insight: Interaction, Slide and Buy are too sparse and volatile; Click and Duration diverge in the long-tail (high-score) region. Only Effective Completion satisfies both principles, making it a robust and cheaply obtainable anchor for content recommendation pipelines.

Conclusion

Uniboost turns abstract model scores into interpretable business values through posterior value alignment, replaces coupled boosting with independent linear boosting, and delivers both micro-level efficiency gains and macro-level strategic guidance via ROI attribution. Fully deployed in Taobao Content Feeds as the core traffic allocation system, serving hundreds of millions of users daily.