Hierarchically Constrained Adaptive Ad Exposure (HCA2E) for Dynamic Feed Advertising

Most feed scenarios present a mixture of organic content and commercial ads. Fixed‑position ad placement is simple but ignores user preferences, leading to low efficiency. This work focuses on dynamic ad exposure, modeling it as a dynamic knapsack problem and proposing the Hierarchically Constrained Adaptive Ad Exposure (HCA2E) method.

HCA2E jointly optimizes platform revenue and user experience under multi‑objective constraints while preserving the incentive‑compatible and individually rational properties of the ad auction. The method achieves near‑optimal platform performance, high computational efficiency, and stable effects.

The problem is modeled by treating each user request as an item in a knapsack, where the value is the incremental benefit of showing an ad and the weight is the number of ads shown. Global constraints limit the overall commercialization rate, while per‑request constraints enforce the highest ad position and minimum spacing between ads.

To solve the dynamic knapsack, a hierarchical optimization framework decouples global traffic selection from per‑request ad placement. A greedy algorithm ranks requests by value‑to‑weight ratio, and a beam‑search based template search finds near‑optimal ad placement patterns under hierarchical constraints.

A preservation‑order (保序) strategy is introduced to keep the relative order of recommendation and ad results consistent with upstream systems, guaranteeing auction properties and reducing the search space from exponential to linear in the number of positions.

Offline experiments on Taobao feed data compare HCA2E with Fixed, Whole‑page Optimization (WPO), and Gap Effect Algorithm (GEA). HCA2E achieves superior Pareto fronts, higher revenue and GMV, and more balanced ad position distribution. Beam size analysis shows diminishing returns beyond a moderate size.

Online A/B tests in multiple feed scenarios (home, collection, cart, checkout, order list, logistics) confirm that HCA2E improves both ad‑side and recommendation‑side metrics, while delivering ads at slightly later positions, consistent with offline findings. Real‑time feedback control stabilizes the commercialization rate with minimal daily fluctuation.

The paper concludes with future directions, including cross‑scenario commercialization allocation and whole‑page evaluation via contextual modeling, and invites AI‑oriented talent to join the Alibaba Mama ad‑tech team.