Using Predicted Weights for Ad Delivery

TL;DR

This paper empirically evaluates a proportional weights algorithm for online ad delivery, demonstrating its superior performance over baselines on Yahoo! data and extending theoretical guarantees to non-stationary capacities and large random-order scenarios.

Contribution

It provides a comprehensive empirical analysis of the algorithm's performance on real data and extends theoretical results to dynamic capacity settings and large-scale models.

Findings

The algorithm outperforms greedy and ranking baselines on Yahoo! ad impression data.

It maintains strong theoretical guarantees even with non-stationary advertiser capacities.

Near-optimal performance is achieved in large random-order models.

Abstract

We study the performance of a proportional weights algorithm for online capacitated bipartite matching modeling the delivery of impression ads. The algorithm uses predictions on the advertiser nodes to match arriving impression nodes fractionally in proportion to the weights of its neighbors. This paper gives a thorough empirical study of the performance of the algorithm on a data-set of ad impressions from Yahoo! and shows its superior performance compared to natural baselines such as a greedy water-filling algorithm and the ranking algorithm. The proportional weights algorithm has recently received interest in the theoretical literature where it was shown to have strong guarantees beyond the worst-case model of algorithms augmented with predictions. We extend these results to the case where the advertisers' capacities are no longer stationary over time. Additionally, we show the algorithm has near optimal performance in the random-order arrival model when the number of impressions and the optimal matching are sufficiently large.