In online advertising, it is crucial to identify an optimal ad asset for diverse users to maximize platform revenue. Many algorithms have been developed for selecting personalized ad assets, and have been applied in industrial systems. However, to fully leverage the performance of these algorithms, advertisers are required to submit an appropriate set of ad assets, which we call an \textit{ad asset portfolio}, to the platforms. This problem, which we call \textit{ad asset portfolio optimization (Ad-POP)}, is a crucial issue for online advertising, but it remains underexplored in the existing literature.
To tackle this problem, we first formulate Ad-POP in the contextual combinatorial bandits, and reduce it to an off-policy learning problem (OPL), which aims to optimize a new policy solely using historical logged data collected by a different policy.
Typical OPL methods can be applied to the Ad-POP problem and categorized into two approaches: the independent approach and the exact approach.
The independent approach constructs an ad asset portfolio by adding ads with the highest value, while the exact approach treats an ad set as a single action and directly optimizes ad asset portfolios. However, these approaches suffer from high variance and high bias, respectively.
To address these challenges, we propose a novel algorithm, named \textbf{\textit{Ad Asset Portfolio Optimization via Meta Information (Ad-POM)}}. Our algorithm decomposes an ad selection policy into a first-stage policy for selecting meta information, such as the size of the portfolio and the degree of diversity, and a second-stage policy for selecting the portfolio given the meta information. Specifically, we propose a new policy gradient estimator to learn the first-stage policy. This method can achieve stable optimization since it applies importance weighting only to meta information.
Our comprehensive experiments on real-world data demonstrate that the proposed method can provide substantial improvements in Ad-POP, where existing methods fail due to the large action space and the interactions among ad assets.