Chen et al. (2021) publish AutoCO, a production-facing framework for selecting among automatically composed advertising creatives when the combinatorial space is large and click feedback is sparse. It is the primary peer-reviewed computational advertising paper indexed here for dynamic creative selection under exploration–exploitation pressure.

AutoCO searches interaction functions between creative elements (beyond factorization-machine inner products), estimates parameter posteriors with stochastic variational inference, and explores with Thompson Sampling. Offline experiments on synthetic and public datasets reduce cumulative regret versus baselines. An online A/B test reports a 7% increase in CTR versus the baseline.

Problem

Smart ad platforms compose creatives from advertiser materials — templates, fonts, colors, image crops. The example given in the paper: 4 templates × 10 fonts × 10 colors × 5 picture sizes = 2,000 creatives per product image. Impression budgets are thin, creatives turn over quickly (the authors cite more than 4 million new creatives per day on one popular placement), and greedy max-CTR selection under high-variance estimates is suboptimal.

Method

  • Interaction search. One-shot AutoML search over element interaction operators (max, min, plus, multiply, concat) instead of assuming FM inner products.
  • Uncertainty. Variational posterior over parameters so CTR estimates carry usable uncertainty under sparse feedback.
  • Exploration. Thompson Sampling draws creatives from that posterior, balancing proven combinations against untested ones.

Reported result

The online A/B test is the field outcome: AutoCO produced a 7% CTR lift against the production baseline. Offline evaluation is framed as cumulative regret against competing bandit and CTR baselines.

Why this is in the Darwin index

This paper is the concrete ACM record behind Darwin’s combinatorial-selection literature note (cross-element assembly). Where Pieters/Wedel measure human attention to creative surfaces, AutoCO measures algorithmic selection among those surfaces under auction sparsity. Both are Pillar 1; neither substitutes for Darwin Labs field datasets.

Source of record: doi:10.1145/3442381.3449909 · ACM DL.