Factorized Digital Component Selection for Private On-Device Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems face challenges in performing full model prediction for content serving on user devices while preserving user privacy, as they lack sufficient computation resources and hardware like Tensor Processing Units (TPUs, leading to less accurate predictions with lightweight models and increased power consumption and latency.
Innovation Solution
Implement factorized models that utilize server-generated and locally generated embeddings on user devices, allowing ad predictions without requiring specific hardware, by using a combination of server and locally generated embeddings, and maintaining user privacy through a trusted execution environment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If full model prediction is used on user devices, then prediction accuracy is improved, but device computation resources are insufficient and hardware requirements increase
Solution Approach 1:
The prediction model is segmented into two parts: server-generated embeddings (ad embeddings, query embeddings) and locally-generated embeddings (user embeddings). The heavy computational parts are moved to the server, while only lightweight factorized models remain on user devices for combining embeddings and generating predictions. This segmentation allows accurate predictions without requiring user devices to run full complex models.
Solution Approach 2:
Embeddings act as an intermediary representation that bridges server-side full models and client-side lightweight models. The server generates comprehensive embeddings from full models, which are then transmitted to user devices where factorized models combine these pre-generated embeddings with local user embeddings to produce predictions. This intermediary approach enables accurate predictions while reducing computational burden on user devices.
2Device complexity
If lightweight models are used on user devices, then device resource requirements are reduced, but model accuracy and features are compromised
Solution Approach 1:
The server performs preliminary action by pre-generating comprehensive embeddings (ad embeddings, query embeddings) using full models before they are needed for prediction. These pre-computed embeddings are then transmitted to user devices, where lightweight factorized models only need to combine them with local user embeddings. This preliminary computation on the server side allows user devices to use lightweight models while still achieving high prediction accuracy through the pre-generated comprehensive embeddings.
3Measurement precision
If user data is transmitted to content platforms, then prediction accuracy is improved, but user privacy is compromised
Solution Approach 1:
The system segments data processing by keeping user data (user embeddings) on the user device and only transmitting non-private contextual information (request data, information about generating user data factor) to content platforms. The content platforms return contextual selection factors based on this limited data, and the user device combines all factors locally to generate predictions. This segmentation maintains prediction accuracy while preserving user privacy by preventing transmission of sensitive user data.
4Reliability
If full model prediction is performed on user devices, then ad serving quality is maintained, but latency and power consumption increase
Solution Approach 1:
The computational workload is segmented between server and client: the server handles computationally intensive embedding generation using full models, while user devices perform only lightweight factorized model operations to combine pre-generated embeddings with local data. This segmentation maintains ad serving quality through comprehensive server-side processing while significantly reducing power consumption on user devices, which only perform simple vector operations rather than running full models.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for digital component selection are described. In one aspect, a method includes receiving, by a trusted computing device and across a trust boundary, digital component selection factors that each correspond to a digital component. The trust boundary defines computing devices which can access private user data without the private user data being transmitted to untrusted third party devices. In response to receiving a request for a digital component, the trusted computing device transmits a non-private contextual request that includes request data and information about generating a user data factor. The trusted computing device receives a contextual response that includes a contextual selection factor that corresponds to the information about generating the user data factor, and selects a digital component based on the corresponding digital component selection factors, the received contextual selection factor, and the user data factor.