Client-Side Ad Selection Using Deep Learning Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Server-side solutions for selecting advertisements are costly and lead to user experience issues due to extended load times, as they require maintaining state data for millions of users and processing ad selections.
Innovation Solution
Implementing deep learning models on client devices to dynamically select advertisements, where a client machine learning model processes metadata to determine relevant ads based on user interactions, reducing reliance on server-side processing and enhancing privacy by maintaining client data locally.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If server-side solutions are used to determine ad selection, then ad selection accuracy can be maintained through centralized data processing, but system cost increases and user experience deteriorates due to extended load times
Solution Approach 1:
The patent divides the centralized server-side ad selection system into distributed client-side machine learning models. Each client device runs its own model locally, segmenting the processing task from the server. This segmentation enables parallel processing across multiple devices, reducing server load and response time while maintaining selection accuracy through locally trained models that leverage user-specific data.
Solution Approach 2:
The patent transitions the ad selection processing from a single dimension (server-side centralized processing) to multiple dimensions by distributing models across numerous client devices. This dimensional shift allows the system to process ad selections locally at each client while still benefiting from centralized model training and updates, thereby reducing latency and improving user experience without sacrificing accuracy.
2Adaptability or versatility
If server-side solutions are used to maintain state data for millions of users, then comprehensive user data can be processed, but system cost and operational complexity increase significantly
Solution Approach 1:
The patent extracts the ad selection processing logic from the server environment and places it directly on client devices through locally executed machine learning models. This extraction eliminates the need for the server to maintain and process state data for millions of users in real-time, significantly reducing system operational complexity and cost while preserving the ability to process comprehensive user data locally.
Solution Approach 2:
Each client device runs its own machine learning model that autonomously processes ad selection decisions using locally stored user data. This self-service approach allows individual devices to make ad selection decisions independently without requiring centralized server intervention for each decision, thereby reducing the operational complexity of maintaining state data for millions of users while preserving comprehensive data processing capabilities.
3Productivity
If client-side deep learning models are implemented, then processing speed improves and costs reduce, but model deployment and maintenance complexity increases
Solution Approach 1:
The patent performs model training and updates in advance on centralized servers, then deploys pre-trained models to client devices. This preliminary action allows the complex model training process to occur once centrally rather than requiring continuous complex deployment operations at each client device, thereby maintaining high processing speed while reducing the ongoing deployment and maintenance complexity.
Data Source
AI summary
Systems, methods, and computer program products to transmit, by a web browser to a web server, a hypertext transfer protocol request for a web page at a first uniform resource identifier (URI). The web browser may receive, from the web server via, the web page and metadata of a plurality of candidate advertisements, the plurality of candidate advertisements determined based on a master machine learning (ML) model. A client ML model executing in the web browser may process the received metadata, the client ML model trained based on prior interactions between one or more users of the web browser and a plurality of previously displayed advertisements. The client ML model may determine based on the processing, a first candidate advertisement of the plurality of candidate advertisements to display in the web browser with the received web page. The web browser may receive, from a second URI, the first candidate advertisement of the plurality of candidate advertisements and output the web page and the first candidate advertisement on a display device.