Privacy-Constrained Conversion Model Training With Mixed Attribution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Increased privacy restrictions hinder the accuracy of conversion probability models in content serving systems due to limited attribution data, making it difficult to train effective models for predicting user interactions with digital content.
Innovation Solution
A system and method that utilizes machine learning models trained with both attributable and un-attributable conversion event information, incorporating a privacy bias, to determine conversion probabilities and select content items for presentation in enhanced privacy environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If privacy restrictions are increased to protect user data, then user privacy is improved, but conversion probability model accuracy deteriorates due to limited attribution data
Solution Approach 1:
The patent segments conversion events into attributable conversion events (with identifiable user actions) and un-attributable conversion events (without direct user action identification). This segmentation allows the system to utilize both types of events for model training while respecting privacy restrictions, thereby maintaining prediction accuracy without compromising user privacy protection.
Solution Approach 2:
The patent applies partial action by using a portion of available conversion data (attributable events) directly, while using another portion (un-attributable events) with modified weighting or processing. This approach allows the model to learn from all available data while appropriately accounting for the reduced reliability of un-attributable events, thus maintaining accuracy under privacy constraints.
2Measurement precision
If only attributable conversion events are used for model training, then model accuracy is improved, but the amount of available training data deteriorates due to privacy restrictions
Solution Approach 1:
The patent introduces an intermediary processing mechanism that handles un-attributable conversion events. Instead of directly using these events as reliable training data, the system processes them through a intermediary layer that applies privacy bias adjustments and weighting schemes. This intermediary approach allows the system to incorporate additional training data while maintaining model accuracy by appropriately accounting for the lower reliability of un-attributable events.
3Quantity of substance
If un-attributable conversion events are incorporated into training data, then the quantity of training data is improved, but data quality deteriorates due to lack of direct user action attribution
Solution Approach 1:
The patent applies parameter changes by modifying how un-attributable conversion events are weighted and processed during model training. Instead of treating all conversion events equally, the system adjusts parameters such as event weighting, loss function coefficients, and training sample selection criteria to account for the lower quality of un-attributable events. This allows the system to incorporate more training data while maintaining overall data quality through parameter adjustments.
Data Source
AI summary
In an example, sets of event information associated with events may be identified. The events may include conversion events. The conversion events may comprise attributed and un-attributed conversion events, as well as aggregated conversion events. Machine learning model training may be performed, using a training set of attributable and un-attributable conversion events and a privacy bias in connection with the un-attributable conversion events, to generate a first machine learning model. Further machine learning model training may be performed, using the aggregated conversion events to generate a second machine learning model. Conversion probabilities associated with content items may be determined using the first and second machine learning models. Attributable content items may be selected for presentation via a client device based upon the conversion probabilities.