Cold-Start Content Population Using Consented and Substitute Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Users joining new services often receive generic content due to regulatory restrictions on sharing user data across organizations, preventing effective personalization based on complete interaction histories.
Innovation Solution
A cold start model (CS model) is trained using user data from multiple sources, incorporating consented data and substitute values for non-consented data to generate personalized content, leveraging a two-tower neural network architecture for content recommendations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If user data from multiple sources is collected to provide personalized content, then personalization quality improves, but regulatory compliance deteriorates due to consent restrictions
Solution Approach 1:
The user data is segmented into multiple sources (first service data, second service data, third service data) with different consent statuses. The system selectively processes data from sources where consent is obtained while handling non-consented sources differently, thereby achieving personalization while complying with regulatory requirements.
Solution Approach 2:
The cold start model acts as an intermediary that processes mixed-quality data (both consented and non-consented) and transforms it into personalized content recommendations. The model handles the complexity of data integration and consent management internally, shielding the user experience from these underlying complexities.
2Measurement precision
If complete user interaction history is required for personalization, then content relevance improves, but system adaptability deteriorates for new users without history
Solution Approach 1:
The system performs preliminary actions by collecting and processing data from multiple service sources before the user actually uses the additional service. The cold start model is trained in advance with diverse data sources, enabling it to generate personalized recommendations immediately when a new user joins, without requiring the user to build up interaction history first.
Solution Approach 2:
The system uses partial data (from sources where consent is obtained) and supplement it with substitute values for missing features from non-consented sources. This partial action approach allows the system to function effectively with incomplete data rather than requiring complete interaction history.
3Loss of information
If substitute values are used for non-consented data, then data completeness improves, but data accuracy deteriorates
Solution Approach 1:
The system prepares substitute values in advance for potential missing features before they are needed. When data from non-consented sources is missing, pre-prepared substitute values are used to cushion the impact of incomplete data, ensuring the model can still function without severe accuracy degradation.
Solution Approach 2:
The system changes the parameter representation by using substitute values for missing features rather than treating them as complete absences. This parameter substitution allows the model to process incomplete data as if it were complete, maintaining functional accuracy even when original data is unavailable.
Data Source
AI summary
Described are systems and processes for generating personalized content for a user upon cold start of an additional service by leveraging first user data having user consent for a first service and providing substitute data for the user for a second service where the user has no second user data or does not provide consent for use of the second user data. The first user data and the substitute data are input to a content model that may be trained in part with some null values. The content model generates personalized content for the additional service. The personalized content may be cached and retrieved, in part, after the user initiates use with the additional service. The personalized content may be presented to the user based on information from the first user data to create a personalized presentation of at least a portion of the cached personalized content.