Optimizing Real-Time Bidding: Machine Learning Approaches for Bid Shading and Winning Price Prediction

Real-Time Bidding (RTB) is a critical mechanism in online advertising, where Demand-Side Platforms (DSPs) must determine optimal bids to maximize returns. The effectiveness of bid shading algorithms heavily depends on accurately forecasting the winning price distribution, which differs significantly between second-price (SPA) and first-price (FPA) auctions. Unlike traditional point-estimation models, winning price prediction faces unique challenges such as data censorship and counterfactual learning, as DSPs often only observe win/loss outcomes rather than true market prices.

To address mixed censorship in auction environments, the authors propose the Multi-task Mixed Censorship Predictor (MMCP). This framework leverages the similarity between first-price and second-price auction environments to transfer market price information. It consists of four modules: feature embedding with mixture-of-experts, a second-price prediction module, a first-second mapping module (FSMM) that aligns distributions using Wasserstein distance, and a first-price prediction module optimized with an Adaptive Censorship Loss (ACL). Experiments on iPinYou and Tencent's Youlianghui datasets demonstrate that MMCP significantly outperforms baseline models in surplus rate.

Furthermore, traditional bid shading suffers from sample selection bias due to over-exploitation, leading models to converge to local optima. To mitigate this, the Cost-Effective active learning for Bid Exploration (CeBE) method is introduced. CeBE balances exploration and exploitation by formulating bid exploration as a stochastic knapsack optimization problem. It quantifies exploration utility using prediction entropy and sample similarity, while calculating exploration cost based on opportunity loss. By applying a greedy algorithm under budget constraints, CeBE effectively gathers representative training data, substantially improving model generalization and advertiser ROI compared to random exploration and standard active learning approaches.

In conclusion, bid shading and winning price prediction remain active research frontiers in computational advertising. Beyond classical distribution estimation and linear programming, modern solutions increasingly incorporate active learning, reinforcement learning, and uplift modeling. The integration of these advanced machine learning techniques enables DSPs to navigate complex auction dynamics, reduce advertising costs, and drive sustainable growth in real-time bidding ecosystems.