Chen et al. (2021) publish AES — Adaptive and Efficient ad creative Selection — for platforms that compose creatives from ingredient trees when CTR feedback is too sparse for naïve enumeration.

AES models compositing ingredients as a tree, uses dynamic programming to select high-CTR creatives, and adapts Thompson sampling over that tree to explore under sparse feedback. On synthetic and real-world datasets, AES outperforms ε-greedy, UCB, LinearUCB, and independent greedy on convergence rate and overall CTR (about +8% CTR vs independent greedy after 200 batches in reported comparisons). Runtime: 165 s vs 730 s full enumeration and 657 s hill climbing on the real dataset.

Problem

Advertising platforms can instantly composite creatives from advertiser materials. The combinatorial space explodes; each creative gets few impressions; CTR estimates have high variance. Exhaustive search is too slow for real-time selection.

Method

  • Ingredient tree. Composition structure enables dynamic programming instead of full creative enumeration.
  • Thompson sampling + DP. Sample posterior CTR parameters, then optimize over the tree for efficient exploration of high-CTR candidates.

Why this is in the Darwin index

AES is the tree/DP companion to AutoCO (WWW ’21), which searches interaction functions and reports a +7% online CTR lift. Together they are the primary computational advertising sources for the combinatorial-selection literature map.

Source of record: doi:10.1609/aaai.v35i5.16516 · AAAI proceedings.