Dynamic Ad Combination Optimization via Explore-Exploit Layer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional recommender systems for advertising struggle to maximize revenue by lacking the ability to dynamically optimize ad combinations based on user segments and contextual factors, leading to suboptimal ad performance.
Innovation Solution
The integration of a thin explore/exploit layer at the frontend ad serving engine, which uses machine learning to generate combination distributions for ads based on predicted performance across different user segments and contexts, allowing for real-time selection of the most effective ad combinations during auctions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional ad selection models are used to select ads based on user information and contextual information, then ad selection can be performed efficiently, but the system cannot dynamically optimize ad combinations to maximize revenue and click-through rates
Solution Approach 1:
The patent segments the ad selection process into multiple independent components: ad selection models for selecting individual ads, and combination models for optimizing combinations of ads. This segmentation allows each component to be trained and optimized independently, managing complexity while enabling dynamic optimization of ad combinations to maximize revenue and click-through rates
Solution Approach 2:
The patent transitions from selecting individual ads in one dimension to selecting combinations of ads across multiple dimensions. The combination models evaluate ads based on multiple factors including user information, contextual information, and interaction between ads in the combination, adding dimensional complexity that enables superior revenue optimization
2Reliability
If ad combinations are dynamically optimized based on user segments and contextual factors, then click-through rates and conversion rates are maximized, but the computational complexity and processing requirements increase
Solution Approach 1:
The patent performs preliminary actions by pre-training ad selection models and combination models offline using historical data. The combination distributions are pre-computed based on predicted performance across different user segments and contexts. During real-time ad serving, the system only needs to sample from pre-computed distributions, significantly reducing online computational complexity while maintaining high ad performance
Solution Approach 2:
The system uses machine learning models that automatically learn optimal ad combination strategies from historical performance data. The models self-optimize by training on past interactions between ads and users, eliminating the need for manual configuration and reducing the burden of managing computational complexity
Data Source
AI summary
The present teaching relates to generating combination distributions for ads. A prediction model is obtained via machine learning with respect to a criterion. Training data are associated with multiple ads each having multiple attributes, and include combinations with recorded performance for each ad. Each combination has multiple assets representing respective attributes of an ad. Using the prediction model, performance of each combination of each ad can be predicted and used for generating combination distributions for the ads. Such generated combination distributions are then sent to an explore/exploit layer (EEL) at a frontend ad serving engine so that it can draw a combination associated with an auction winning ad for rendering on a webpage viewed by a user on a user device.