Reciprocity in Machine Learning

TL;DR

This paper investigates the reciprocity of influence and benefit in machine learning models, proposing measures to quantify outflows, inflows, and reciprocity, with initial theoretical and empirical findings suggesting some models are approximately reciprocal.

Contribution

It introduces new measures for assessing influence and reciprocity in machine learning, extending previous influence measures with theoretical and empirical analysis.

Findings

Some classes of models are approximately reciprocal under certain assumptions

Proposed measures effectively quantify influence flows and reciprocity

Open directions for further research are identified

Abstract

Machine learning is pervasive. It powers recommender systems such as Spotify, Instagram and YouTube, and health-care systems via models that predict sleep patterns, or the risk of disease. Individuals contribute data to these models and benefit from them. Are these contributions (outflows of influence) and benefits (inflows of influence) reciprocal? We propose measures of outflows, inflows and reciprocity building on previously proposed measures of training data influence. Our initial theoretical and empirical results indicate that under certain distributional assumptions, some classes of models are approximately reciprocal. We conclude with several open directions.