Transfer Machine Learning for Cookie-Free User Attribute Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models rely on third-party cookies for user attribute prediction, which are being phased out, leading to challenges in accurately predicting user attributes with limited data availability and increased computational and network resource usage.
Innovation Solution
Implementing transfer machine learning techniques to train models using data from subscribed users' pipelines, adapting them for predicting user attributes in environments with limited data, and selecting digital components based on predicted attributes without relying on third-party cookies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If third-party cookies are used for user attribute prediction, then prediction accuracy is maintained, but data privacy concerns increase and cookie availability decreases
Solution Approach 1:
The patent introduces a transfer learning model as an intermediary that bridges the gap between limited first-party data and the need for accurate predictions. The model transfers knowledge from a source domain (subscribed users with rich data) to a target domain (non-subscribed users with limited data), enabling accurate attribute prediction without relying on third-party cookies
Solution Approach 2:
The system changes the parameters of the machine learning approach by shifting from direct third-party cookie-based prediction to transfer learning with feature adaptation. This involves changing how features are extracted and how the model is trained, moving from cookie-dependent features to contextual feature representations that work with limited data
2Productivity
If traditional machine learning models are trained with limited data, then training speed increases, but prediction accuracy deteriorates
Solution Approach 1:
The system performs preliminary action by pre-training the machine learning model on a source domain with abundant data (subscribed users) before adapting it to the target domain. This pre-training phase captures general patterns and knowledge that can be transferred, so that when the model is later applied to non-subscribed users with limited data, it already possesses substantial predictive capability
Solution Approach 2:
The patent adds another dimension to the learning process by introducing a transfer learning framework that operates across two domains (source and target). This dimensional expansion allows the system to leverage knowledge from the source domain to compensate for data scarcity in the target domain, effectively solving the accuracy-speed tradeoff
3Measurement precision
If more data is collected for model training, then prediction accuracy improves, but computational resource usage increases
Solution Approach 1:
The system extracts only the essential knowledge and patterns from the source domain data that are necessary for making accurate predictions in the target domain. Rather than using all available source data continuously, the transfer learning process extracts transferable features and model parameters, reducing the computational burden while maintaining prediction accuracy
4Object-affected harmful factors
If third-party cookies are phased out, then data privacy is improved, but user attribute prediction capability deteriorates
Solution Approach 1:
The system enables self-service by using first-party data from subscribed users to train transfer learning models that can then serve non-subscribed users. This self-contained approach eliminates dependence on third-party cookies while maintaining prediction capability, as the model learns from available first-party data and generalizes to other users
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for using transfer machine learning to predict attributes are described. In one aspect, a method includes receiving, from a client device of a user, a digital component request that includes at least input contextual information for a display environment in which a selected digital component will be displayed. The contextual information is converted into input data that includes input feature values for a transfer machine learning model trained to output predictions of user attributes of users based on feature values for features representing display environments. The transfer machine learning model is trained using training data for subscriber users obtained from a data pipeline associated with electronic resources to which the subscriber users are subscribed and adapted to predict user attributes of non-subscribing users viewing electronic resources to which the non-subscribing users are not subscribed.