Personalized Automated Bidding Framework (PerBid) for Fairness‑Aware Online Advertising

Abstract

This work proposes PerBid, a personalized automated bidding framework that extends a single unified bidding agent into multiple context‑aware agents, each matched to a specific advertiser cluster. An Ad Campaign Profiling Network models the dynamic advertising environment, and clustering groups advertisers with similar contexts. Experiments on Alibaba’s display‑ad platform demonstrate significant improvements in overall performance and fairness.

Background

Uniform auto‑bidding strategies, often based on reinforcement learning, suffer from large performance gaps across heterogeneous advertisers due to differing traffic distributions, budget constraints, and state‑space imbalance. These gaps motivate a personalized approach that can perceive contextual differences.

Preliminary Knowledge

Automated Bidding Strategies

Online ad placement is modeled as a constrained optimization problem. Existing solutions use feedback control or RL to adjust bidding parameters in real time, but a single policy cannot capture the diverse environments of all advertisers.

Personalized Bidding Framework

Ad Campaign Profiling

A profiling network combines static plan features (e.g., target audience, schedule) with dynamic bidding‑log features. Static features are embedded, while dynamic features are encoded per time slot and processed by a GRU to capture temporal evolution, producing a campaign embedding that represents the current environment.

Context‑Aware Bidding Strategy

The campaign embedding is injected as context into the RL state, enabling each agent to make bidding decisions that reflect both real‑time metrics and higher‑level environment information without exploding the state space.

Candidate Strategy Generation

Advertisers are clustered based on their profiling embeddings. For each cluster, a dedicated RL bidding agent is trained. An iterative process alternates between training agents and re‑assigning advertisers to clusters based on performance, yielding a set of specialized policies.

Strategy Matching and Adaptation

For a new campaign, historical data are used to simulate each candidate policy offline; the policy with the best average performance is selected. Cold‑start campaigns receive a default universal policy. Local adaptation fine‑tunes the chosen policy on the campaign’s own data.

Experimental Results

Offline Experiments

Using a dataset of >3,500 campaigns, PerBid outperforms baselines (M‑PID, DRLB, USCB) on both overall ROI and fairness metrics (Generalized Gini, Gini coefficient). The framework reduces the long‑tail effect and improves performance for the bottom 30% of campaigns.

Online A/B Test

Deployed on Alibaba’s platform, PerBid achieved up to 10.85% lift in key metrics compared with the production USCB algorithm, confirming its effectiveness in a real‑world setting.

Conclusion

PerBid addresses the effectiveness disparity caused by a single unified bidding policy by introducing advertiser‑specific, context‑aware agents. Comprehensive offline and online evaluations validate its ability to boost average performance while enhancing fairness across heterogeneous advertisers.