Recommender System Segmentation for Long-Term Engagement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommender systems are optimized for short-term outcomes and fail to effectively promote long-term user engagement and retention, as they do not adequately consider repeat consumption or retention behaviors, and face challenges with measurement, attribution, and coordination in reinforcement learning approaches.
Innovation Solution
The development of a system that trains separate machine learning models for short-term and long-term outcomes, combining their outputs to rank content items based on expected future engagement, using a surface-independent long-term recommender system and surface-specific short-term recommenders, to drive longer-term user habits and behaviors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If recommendation algorithms are optimized for short-term outcomes (maximizing clicks or 30-second streams), then immediate interaction with content is driven, but long-term behaviors and habits (repeat consumption or retention) are not considered
Solution Approach 1:
The patent segments the recommendation system into multiple independent algorithms, each optimized for different time horizons (e.g., one algorithm for immediate clicks, another for 7-day retention, another for 30-day retention). These segmented algorithms operate in parallel and their outputs are combined to produce the final recommendation ranking, allowing short-term and long-term goals to be pursued simultaneously without one compromising the other.
Solution Approach 2:
The patent merges the outputs of multiple recommendation algorithms with different optimization objectives into a unified recommendation list. By combining the rankings or scores from short-term optimized algorithms and long-term optimized algorithms, the system achieves a balanced recommendation strategy that drives both immediate engagement and long-term user retention.
2Reliability
If reinforcement learning approaches are used to address long-term optimization, then long-term goals can be optimized, but significant challenges with measurement, attribution, and coordination arise
Solution Approach 1:
The patent segments the complex reinforcement learning problem into smaller, more manageable components by creating separate algorithms for different time horizons and objectives. This segmentation reduces the measurement and attribution complexity by isolating specific objectives (e.g., 7-day retention vs. 30-day retention) that can be measured and optimized independently, rather than attempting to measure and optimize all long-term outcomes simultaneously in a single complex RL system.
3Device complexity
If a single recommendation system is used to optimize both short-term and long-term outcomes, then coordination is simplified, but the system cannot effectively promote long-term user engagement and retention
Solution Approach 1:
The patent divides the recommendation system into multiple specialized algorithms, each responsible for optimizing specific objectives (immediate clicks, 7-day retention, 30-day retention). This segmentation allows each algorithm to focus on its specific objective without being compromised by conflicting objectives, thereby effectively promoting long-term user engagement while maintaining a manageable system architecture through modular design.
Solution Approach 2:
The patent combines the outputs of multiple specialized recommendation algorithms into a unified recommendation list. By merging the rankings from algorithms optimized for different time horizons, the system achieves effective long-term engagement promotion while keeping the overall architecture relatively simple through the use of standardized combination mechanisms (e.g., weighted scoring or rank aggregation).
Data Source
AI summary
The present application describes various methods and devices for providing content to users. In one aspect, a method includes, for each content item of a set of content items, obtaining a score for the content item using a recommender system, the score corresponding to a calculation of subsequent repeated engagement by a user with the content item. The method also includes ranking the set of content items based on the respective scores and providing recommendation information to the user for one or more highest ranked content items in the set of content items.