Online Ranking Re-Ranking for Fairness Under Concept Drift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning ranking systems face challenges in maintaining fairness constraints while maximizing utility, particularly in online environments where data distributions can change over time, leading to violations of fairness constraints in non-training data.
Innovation Solution
Implement online post-processing techniques that involve a sequence of ranking models, including an initial trained model and a re-ranking model, to ensure that fairness constraints are satisfied while optimizing cumulative utility. This includes deterministic policies like Fair Queues and Greedy Fair Swap, as well as learned policies through Locally Optimal Learning to Search with Queues (L2SQ), to manage unknown distributions and concept drift.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a trained ranking model is used to maximize utility, then cumulative utility is improved, but fairness constraints are violated due to unknown data distributions and concept drift
Solution Approach 1:
The patent introduces a re-ranking model as an intermediary component between the initial ranking model and the final output. The re-ranking model takes the initial rankings and adjusts them to satisfy fairness constraints while preserving utility. This mediator resolves the contradiction by decoupling the utility maximization function (initial model) from the fairness enforcement function (re-ranking model), allowing both objectives to be achieved simultaneously.
Solution Approach 2:
The system performs preliminary actions by maintaining a buffer of unranked items and pre-computing fairness metrics before final ranking decisions are made. This allows the system to proactively adjust rankings to prevent fairness violations rather than reactively correcting them, ensuring fairness constraints are satisfied while maximizing cumulative utility over time.
2Reliability
If online post-processing is implemented to satisfy fairness constraints, then fairness is improved, but system complexity increases due to multiple ranking models and re-ranking operations
Solution Approach 1:
The patent segments the ranking system into distinct functional components: an initial ranking model for utility maximization, a re-ranking model for fairness enforcement, and a buffer management system. This segmentation allows each component to specialize in its function, making the overall complex system more manageable and maintainable while achieving both utility and fairness objectives.
Solution Approach 2:
The re-ranking model serves as an intermediary layer that simplifies the complexity by providing a unified interface for fairness enforcement. Rather than embedding complex fairness logic throughout the entire ranking pipeline, the intermediary re-ranking model centralizes fairness constraints, making the system easier to control and adjust.
3Reliability
If re-ranking is performed considering previous batches, then fairness across batches is improved, but processing time increases due to online post-processing requirements
Solution Approach 1:
The patent maintains continuity of useful action by processing batches continuously through the ranking and re-ranking pipeline without interrupting the data flow. The buffer of unranked items enables smooth transitions between batches, allowing fairness adjustments to be made incrementally as new data arrives, rather than requiring batch-wise reprocessing that would cause time delays.
Solution Approach 2:
The system performs preliminary fairness assessments and maintains running statistics of fairness metrics across batches before final ranking decisions. This preliminary action allows the re-ranking model to make quick adjustments based on pre-computed fairness states, reducing the actual processing time required for each batch while still achieving inter-batch fairness.
Data Source
AI summary
Online post-processing may be performed for rankings generated with constrained utility maximization. A stream of data items may be received. A batch of data items from the stream may be ranked according to a ranking model trained to rank data items in a descending order of relevance. The batch of data items may be associated with a current time step. A re-ranking model may be applied to generate a re-ranking of the batch of data items according to a re-ranking policy that considers the current batch and previous batches with regard to a ranking constraint. The re-ranked items may then be sent to an application.