t in git g

Provider-side fairness for recommender systems using Fairness Maximal Marginal Relevance

By Dora Jambor

My submission to NeurIPS'18 "ML for the Developing World (ML4D)" workshop.

This paper demonstrates a provider-side fairness extension of Fairness Maximal Marginal Relevance (FMMR) technique described by Karako and Manggala (2018). This extension is demonstrated by a case study where, based on the geographic location of various item providers, the objective is to bolster local businesses in small countries competing against larger providers in a global online marketplace. Competitiveness is measured as the exposure of item providers in recommendations. It is shown that standard relevance-based recommendations often provide more exposure to providers that are larger in size with more user interactions. FMMR is used as a post-processing technique to tune the interplay between standard relevance-based recommender systems and fairness of item provider exposure. Using contextual embeddings to represent item providers, FMMR is able to mitigate the large differences in provider exposures that standard recommenders most often do not account for.


Chen Karako and Putra Manggala. 2018. Using Image Fairness Representations in Diversity-Based Re-ranking for Recommendations. In Adjunct Publication of the 26th Conference on UserModeling, Adaptation and Personalization. ACM, 23–28.

Handcrafted in New York City, design & code by
Dora Jambor