Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving Minimax Group Fairness in Sequential Recommendation

Krishna Acharya, David Wardrope, Timos Korres, Aleksandr Petrov, Anders Uhrenholt

TL;DR

In experiments on two real-world datasets, it is shown that the DRO methods outperform standard training, with CVaR delivering the best results, and that Group and Streaming DRO are sensitive to the choice of group used for loss computation.

Abstract

Training sequential recommenders such as SASRec with uniform sample weights achieves good overall performance but can fall short on specific user groups. One such example is popularity bias, where mainstream users receive better recommendations than niche content viewers. To improve recommendation quality across diverse user groups, we explore three Distributionally Robust Optimization(DRO) methods: Group DRO, Streaming DRO, and Conditional Value at Risk (CVaR) DRO. While Group and Streaming DRO rely on group annotations and struggle with users belonging to multiple groups, CVaR does not require such annotations and can naturally handle overlapping groups. In experiments on two real-world datasets, we show that the DRO methods outperform standard training, with CVaR delivering the best results. Additionally, we find that Group and Streaming DRO are sensitive to the choice of group used for loss computation. Our contributions include (i) a novel application of CVaR to recommenders, (ii) showing that the DRO methods improve group metrics as well as overall performance, and (iii) demonstrating CVaR's effectiveness in the practical scenario of intersecting user groups.

Improving Minimax Group Fairness in Sequential Recommendation

TL;DR

In experiments on two real-world datasets, it is shown that the DRO methods outperform standard training, with CVaR delivering the best results, and that Group and Streaming DRO are sensitive to the choice of group used for loss computation.

Abstract

Training sequential recommenders such as SASRec with uniform sample weights achieves good overall performance but can fall short on specific user groups. One such example is popularity bias, where mainstream users receive better recommendations than niche content viewers. To improve recommendation quality across diverse user groups, we explore three Distributionally Robust Optimization(DRO) methods: Group DRO, Streaming DRO, and Conditional Value at Risk (CVaR) DRO. While Group and Streaming DRO rely on group annotations and struggle with users belonging to multiple groups, CVaR does not require such annotations and can naturally handle overlapping groups. In experiments on two real-world datasets, we show that the DRO methods outperform standard training, with CVaR delivering the best results. Additionally, we find that Group and Streaming DRO are sensitive to the choice of group used for loss computation. Our contributions include (i) a novel application of CVaR to recommenders, (ii) showing that the DRO methods improve group metrics as well as overall performance, and (iii) demonstrating CVaR's effectiveness in the practical scenario of intersecting user groups.

Paper Structure

This paper contains 27 sections, 6 equations, 2 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Sequential recommenders and training methods
  3. Sequential recommenders
  4. Standard training (ERM)
  5. Cost-sensitive losses
  6. Class balanced (CB) loss:
  7. Inverse propensity weighted (IPW) loss:
  8. Distributionally robust optimization (DRO):
  9. Conditional Value at Risk (CVaR) DRO:
  10. Group and Streaming DRO:
  11. Loss with intersecting groups:
  12. Experiments
  13. Experimental setup
  14. Datasets and preprocessing:
  15. User groups:
  16. ...and 12 more sections

Figures (2)

  • Figure 1: Percentage increase in NDCG@20 using DRO methods relative to standard training on the Retailrocket dataset: (i) increase for niche, diverse and popular groups (ii) increase for short, medium, and long sequence users.
  • Figure 2: Percentage increase in NDCG@20 using DRO methods relative to standard training on the Movielens-1M dataset: (i) increase for niche, diverse, and popular groups, and (ii) increase for short, medium, and long sequence users.