Optimizing Xiaohongshu Splash Screen Ads: Flow Selection and Dynamic Decision Mechanisms

Xiaohongshu’s splash-screen ads serve as a key brand‑advertising channel, requiring a balance between guaranteed delivery volume and advertising effectiveness while preserving user experience.

The platform introduced a “traffic‑optimal + dynamic decision” solution. First, the traffic‑optimal component formalizes the allocation problem, derives an optimal distribution strategy, and implements online traffic allocation through feedback regulation.

Second, to keep user‑perceived startup time low, a dynamic decision mechanism was designed. Because the app launch window is too short for a full request‑decision‑response cycle, the system adopts an asynchronous approach: the decision for the next splash exposure is prepared in advance and cached locally, allowing immediate display on the next launch.

Problem Formalization : The whole‑day traffic set, order set, per‑order volume guarantees, predicted CTR (pCTR), and order‑specific click‑value weights are defined. The allocation is modeled as a linear‑programming (LP) problem with volume‑guarantee constraints and exclusive‑traffic constraints. When the LP is infeasible due to over‑selling, a penalty weight for shortage is introduced, converting the problem into a solvable form.

The dual‑derived optimal rank‑score formula determines which order wins each traffic unit. If all rank scores are negative, no ad is shown.

Online Allocation Strategy : Assuming known total traffic, the LP yields optimal order weights, which are then adjusted in real time via feedback based on each order’s delivery progress, dynamically updating the allocation.

Asynchronous Decision Mechanism : Ads are pre‑loaded, and the decision for the next exposure is computed and cached after the current exposure. This avoids the startup‑time bottleneck.

Dynamic Decision Mechanism : To mitigate the latency of asynchronous decisions, the system increases the frequency of asynchronous requests and applies real‑time decisions for “cold‑start” launches (app started from scratch). For “hot‑start” launches (app resumed from background), cached decisions are used.

The combined “traffic‑optimal + dynamic decision” approach significantly improves click‑through rate (CTR) while meeting volume guarantees and preserving the short startup experience.

Future work includes more sophisticated parameter control (e.g., reinforcement learning or model predictive control), incorporating user‑behavior prediction into decision latency reduction, and exploring real‑time material loading to reduce client storage overhead.