Incremental Bidding Optimization via Bid Randomization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current uplift models for electronic advertisements are costly, inefficient, and do not directly inform bidding or recommendation strategies, as they require randomized control trials that incur processing overhead and revenue loss, and they focus on individual objectives rather than uplift, limiting their ability to improve incrementality effectively.
Innovation Solution
The system employs a method of bid randomization using a statistical technique called importance sampling to simulate different bidding policies without affecting production environments, allowing for efficient data collection and optimization of bidding policies to improve incrementality by multiplying outgoing bids with a random factor and capturing data on user interactions, thereby estimating the best bidding policy within a family of policies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If randomized control trials are used to learn uplift models, then incrementality can be measured, but processing overhead and revenue loss increase
Solution Approach 1:
The patent creates a synthetic control group by copying and augmenting treatment group data through bid randomization. Instead of requiring a separate control group that misses ad opportunities, the system generates counterfactual outcomes by randomizing bids within the treatment group, eliminating revenue loss while maintaining measurement capability
Solution Approach 2:
The system performs bid randomization and data collection during the normal treatment phase before any formal evaluation. By embedding the experimental mechanism within routine operations, the system collects necessary data without additional processing overhead or revenue opportunity loss
2Reliability
If separate models are learned for control and exposed populations, then each population's objective is optimized, but uplift focus is lost and capacity is wasted
Solution Approach 1:
The patent merges the control and treatment group modeling into a single unified model that directly optimizes for uplift. By combining the datasets and using bid randomization to create synthetic control outcomes, the system learns one model that captures the differential impact on conversions, eliminating the need for separate models and their associated complexity
Solution Approach 2:
The unified model serves multiple functions simultaneously: it learns from both treatment and control outcomes, optimizes for uplift directly, and can be applied to any user population. This multi-functional approach replaces the need for separate population-specific models while maintaining comprehensive coverage
3Measurement precision
If only the first interaction of users is used for learning, then selection bias is avoided, but available data shrinks over time
Solution Approach 1:
The system implements periodic bid randomization at multiple interaction points throughout the user journey. By systematically applying randomization at regular intervals (first interaction, second interaction, etc.), the system captures data from multiple time points while maintaining experimental validity, thereby expanding the available data volume
Solution Approach 2:
The system dynamically adjusts the randomization strategy based on user interaction history. Instead of static first-interaction-only sampling, the system adaptively applies bid randomization across multiple interactions, allowing the data collection mechanism to evolve and capture richer information while maintaining causal inference integrity
4Measurement precision
If uplift models are used, then incrementality can be inferred, but the bidding or recommendation strategy is not directly informed
Solution Approach 1:
The system implements a direct feedback loop where bid randomization results immediately inform bidding strategy optimization. The learned uplift model feeds back into the bidding system, automatically adjusting bids based on predicted incremental impact, making the connection between measurement and action explicit and operational
Solution Approach 2:
The patent introduces an intermediary optimization layer that translates uplift model predictions into actionable bidding strategies. This intermediary component automatically converts incremental quality inferences into bid adjustments, eliminating the need for manual strategy development and ensuring direct alignment between measurement and execution
Data Source
AI summary
Methods and systems are described herein for incremental bidding for electronic advertisements. A computing device generates, for a user during a first time period, first randomized bids for available impression opportunities, the first randomized bids based upon an estimated value and a first random factor and using a context of bid requests. The computing device transmits the first randomized bids to a remote device. The computing device determines whether interaction events associated with impression opportunities occurred. The computing device estimates second randomized bids based upon an estimated value and a second random factor and using the context. The computing device estimates a relative incrementality of a bidding policy based upon the second randomized bids over a bidding policy based upon the first randomized bids, and optimizes the bidding policy based upon second randomized bids. The computing device changes a production bidding policy based upon the optimized policy.