Ad Preference Embedding Model for Lookalike Audience Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing creative generation and trafficking systems are inefficient, requiring significant effort and cost, and lack automation capabilities, especially in generating targeted creative content based on minimal input signals, and struggle with user ad preference modeling due to data sparsity and the 'cold-start' problem.
Innovation Solution
A computer-implemented method and system that automates creative generation and trafficking by using a signal-label model to map activity data points and labels, and employs Locality-Sensitive Hashing to find similar users based on ad preferences, enabling efficient creative content processing and user similarity analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If manual processes are used for creative generation and trafficking, then flexibility and control are maintained, but productivity and speed are significantly reduced
Solution Approach 1:
The system enables self-service automation through unsupervised learning models that automatically generate creative content, perform A/B testing, and optimize trafficking strategies without requiring manual intervention at each step, while still allowing human oversight when needed
Solution Approach 2:
The system performs preliminary actions by pre-generating multiple creative variations, pre-configuring trafficking parameters, and pre-conducting A/B tests before full-scale deployment, allowing faster response times when market conditions change
2Measurement precision
If comprehensive user data collection is implemented to improve ad preference modeling, then measurement precision improves, but user privacy concerns and data complexity increase
Solution Approach 1:
The system extracts only the most relevant features from user data for ad preference modeling, using unsupervised learning to identify key patterns without requiring comprehensive data collection, thereby maintaining accuracy while reducing data complexity and privacy concerns
Solution Approach 2:
The system transforms raw user data into compressed latent representations through neural network embeddings, changing the parameter space from high-dimensional raw data to lower-dimensional meaningful features that capture ad preferences more efficiently
3Adaptability or versatility
If multiple tools and interfaces are used for creative generation, then functionality and versatility are improved, but ease of operation deteriorates due to complex workflows
Solution Approach 1:
The system implements a universal platform that handles multiple creative types (images, video, audio, interactive) and trafficking functions through a single interface, using modular architecture where the same core components serve multiple purposes across different creative formats
Solution Approach 2:
The system merges previously separate functions for creative generation, A/B testing, and trafficking optimization into an integrated automated workflow, allowing users to manage all aspects through a unified system rather than switching between multiple tools
4Extent of automation
If extensive engineering efforts are invested in automated workflows, then extent of automation improves, but device complexity and development time increase
Solution Approach 1:
The system uses template-based approaches where successful creative patterns and trafficking strategies are captured as reusable templates, allowing automated workflows to replicate proven approaches without requiring complex custom engineering for each new campaign
Data Source
AI summary
Methods, systems and computer program products for automating the association of messages. Data points associated with at least one client device associated with an identifier are logged into an activity database. Labels corresponding to message records are retrieved. Message-signal values representing behavior associated with at least a subset of the message records are also retrieved. The labels are merged with the message-signal values to generate a signal-label collection. A signal-label model is trained based on the signal-label collection, thereby generating a trained signal-label model. A mapping of the one or more activity data points and the plurality of labels are then generated. The embedding that is generated can then be used to find custom audiences.