Hierarchical Bayesian Framework for Ad CTR Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current click-through rate (CTR) prediction technologies in online advertising struggle to accurately integrate information from multiple dimensions, such as users, publishers, and advertisers, leading to inconsistent and unreliable revenue estimates due to high variance and skewed data distributions.
Innovation Solution
A dynamic hierarchical Bayesian framework is implemented, using tensor decomposition to construct an integrated framework that performs inference across all dimensions, allowing key information to be shared and enabling accurate CTR prediction by determining marginal prior and posterior probabilities for each party hierarchy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If separate machine learning models are used for each dimension (Advertiser, Publisher, User), then model complexity is reduced and ease of operation is improved, but measurement precision and reliability of CTR prediction deteriorate due to inability to integrate multi-dimensional information
Solution Approach 1:
The patent combines separate machine learning models for Advertiser, Publisher, and User dimensions into a single integrated ensemble model. This ensemble model processes features from all three dimensions simultaneously and generates unified CTR predictions, thereby improving measurement precision while maintaining operational feasibility through modular architecture.
Solution Approach 2:
The patent segments the CTR prediction problem into three distinct dimension hierarchies (Advertiser, Publisher, User), each with its own feature set and model components. By segmenting the input features and processing them through specialized sub-models within the ensemble, the system achieves both operational simplicity and prediction accuracy.
2Measurement precision
If an integrated multi-dimensional framework is created to consistently integrate information from all dimensions, then measurement precision and reliability of CTR prediction are improved, but device complexity and difficulty of detecting and measuring increase
Solution Approach 1:
The framework is segmented into three independent dimension hierarchies (Advertiser, Publisher, User), each with its own feature extraction and modeling components. This segmentation allows the complex integrated framework to be built from manageable modular units, reducing overall structural complexity while maintaining integration benefits.
Solution Approach 2:
The patent creates a universal ensemble model structure that can process features from multiple dimensions simultaneously. This multi-functional framework handles Advertiser, Publisher, and User features through a unified architecture, reducing complexity compared to creating separate specialized models for each dimension.
3Ease of operation
If traditional statistical models are used for CTR prediction, then ease of operation is maintained, but measurement precision deteriorates due to high variance and skewed data distributions in click data
Solution Approach 1:
The patent changes the modeling approach from traditional statistical methods to machine learning-based ensemble models. This parameter change in the algorithmic approach allows the system to better handle high variance and skewed data distributions by using techniques such as feature engineering, regularization, and ensemble averaging, thereby improving measurement precision while maintaining operational feasibility.
Data Source
AI summary
The present disclosure relates to a computer system configured establish and utilize a database for online ad realization prediction in an ad display platform associated with N parties, wherein N is a positive integral greater than 1. The computer system is configured obtain a party hierarchy for each of the N parties including a plurality of features of the party; select a target ad display event including N features, each of the N features corresponding to a node in a party hierarchy; obtain a prior probability reflecting an unconditional probability of ad realization occurrence at the target ad display event among all possible ad display events; for each of the N features: determine a marginal prior probability by decomposing components associated with the other N−1 features from the prior probability; determine a marginal posterior probability based on the marginal prior probability; and save the marginal posterior probability in the corresponding node of the party hierarchy.