Bayesian Conversion Rate Estimation for Low-Volume Ad Terms
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional online advertising systems inaccurately calculate bid amounts for low-volume terms due to lack of statistically significant information, leading to imprecise conversion rate estimates.
Innovation Solution
A computer-based method that computes an estimated conversion rate for low-volume terms by analyzing statistics from related terms, using distribution functions to represent the distribution of conversion events and probabilities, and combining these to generate a posterior distribution for accurate bid calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional systems use average conversion rates from high-volume terms for low-volume terms, then bid calculations become simpler, but precision and accuracy of bid estimates deteriorate
Solution Approach 1:
The patent introduces a Bayesian statistical model as an intermediary between observed conversion data and bid calculations. This model uses prior distributions and likelihood functions to combine limited observed data with prior knowledge, producing more accurate posterior estimates for low-volume terms without requiring complex manual adjustments to each bid calculation
Solution Approach 2:
The patent transforms the conversion rate estimation problem by changing from direct observation to probabilistic inference. It introduces parameters such as prior mean, prior variance, and posterior distributions to represent uncertainty, allowing the system to handle low-volume terms with statistically significant estimates rather than simple averages
2Measurement precision
If systems collect more data for low-volume terms to improve statistical significance, then conversion rate accuracy improves, but time and computational resources increase
Solution Approach 1:
The patent performs preliminary action by establishing prior distributions based on historical data and domain knowledge before new data arrives. This allows the system to make informed estimates immediately rather than waiting for sufficient data to accumulate, reducing the time loss while maintaining statistical rigor
Solution Approach 2:
The Bayesian framework provides continuous feedback as new conversion data arrives, updating the posterior distribution iteratively. This allows the system to improve accuracy progressively without requiring large batches of data collection, reducing overall time loss while achieving statistical significance
3Productivity
If systems ignore low-volume term conversion rates and use only high-volume term averages, then computational resources are saved, but bid precision for low-volume terms deteriorates
Solution Approach 1:
The patent merges information from multiple sources: observed conversion data for low-volume terms, prior distributions from historical data, and domain knowledge. This combination produces accurate bid estimates for low-volume terms while maintaining processing efficiency through the mathematical framework of Bayesian inference
Data Source
AI summary
An estimated conversion rate for a desired advertisement term is calculated. A total number of conversion events for terms having a conversion rate is determined for each of a plurality of conversion rates, and data is generated that describes a first distribution function representing the conversion rates and the associated numbers of conversion rates. Additionally, data describing a second distribution function representing the probability of a given number of conversion events occurring is also generated. Based at least in part on the first distribution function and the second distribution function, the estimated conversion rate for the desired advertisement term is computed.