Contextual Bandit Ad Selection via Embedding Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for selecting ads using Thompson Sampling are slow to learn customer online shopping behaviors and fail to adapt quickly to preference shifts in non-stationary environments, necessitating a more responsive and robust approach that leverages embedding vectors, search queries, page taxonomy, and strong model signals like historical CTR.
Innovation Solution
A system and method that utilize a computing device to assign parameter sets to marketing ads, filter and select ads based on query features, determine context feature sets, and optimize these sets over time using reward metrics, employing a contextual bandit approach to maximize ad relevance and revenue.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If Thompson Sampling approach is used for ad selection, then the system can learn customer behaviors, but it takes long time to adapt to preference shifts in non-stationary environments
Solution Approach 1:
The system pre-processes and stores embedding vectors for ads, queries, and page taxonomy in advance using deep learning models. This preliminary action enables rapid retrieval and comparison when ads need to be selected, eliminating the need for time-consuming real-time learning while maintaining adaptability to preference shifts through pre-computed representations.
Solution Approach 2:
The patent replaces the traditional Thompson Sampling mechanical approach with a deep learning-based semantic matching system. Instead of relying on iterative statistical sampling, the system uses embedding vectors and neural network models to directly compute ad relevance, achieving both speed and adaptability through intelligent algorithms rather than mechanical trial-and-error processes.
2Productivity
If traditional ad selection methods are used, then the system structure remains simple, but the system fails to leverage embedding vectors, search queries, page taxonomy, and model signals for optimized ad selection
Solution Approach 1:
The patent creates a universal ad selection system that integrates multiple data sources (embedding vectors, search queries, page taxonomy, model signals) into a unified framework. The deep learning models process diverse inputs simultaneously to generate comprehensive ad relevance scores, achieving high productivity through multi-functional processing rather than separate specialized systems.
Solution Approach 2:
The system introduces embedding vectors as intermediary representations that bridge different data sources (ads, queries, page taxonomy). These vectors serve as a common language that enables seamless integration and comparison of heterogeneous information, allowing the complex system to function cohesively through standardized intermediate representations.
Data Source
AI summary
The disclosed subject matter relates to a system and method for selecting/recommending ads based on a contextual bandit approach. The disclosed approach leverages various embedding vectors of item, search, page taxonomy trained based on traffic data via advanced deep learning models, and uses model signals (e.g. historical CTR, item price, rating, quality) from other ad placements. The learning mechanism on top of the current methodology to automatic chooses the best feature sets and adjust model performance over time. The contextual bandit model performs better with respect to CTR than the Thompson Sampling model, and achieves lower regret and faster convergence over time.