Fairness and Diversity in Social-Based Recommender Systems

Dimitris Sacharidis, Carine Pierrette Mukamakuza, Hannes Werthner
E-Commerce Research Unit, TU Wien, Vienna, Austria

In social networks, the phenomena of homophily and influence explain the fact that friends tend to be similar. Social-based recommenders exploit this observation by incorporating the social structure in collaborative filtering techniques. In practice, these recommenders tend to make friends appear more similar compared to non-socially aware techniques. Various proposals have demonstrated the benefit of incorporating social connections. But at what cost? In this work, we show that there exist users that are mistreated in social recommenders. Specifically, their individual preferences are suppressed more compared to other users in their social circle. We seek to identify who they are and develop techniques that protect them, without severely affecting the effectiveness of the recommender.

Social-Based Recommender Systems

  • the mechanisms of homophily and social influence suggest we share preferences and tastes with our friends
  • social-based recommenders exploit this phenomenon
    • to improve recommendation accuracy,
    • to increase coverage,
    • to handle cold-start users
  • a key idea is social regularization, which enforces constraints on the learned latent features of users (user embeddings)
    • if two users are socially close, they should have similar latent features

Impact of Social Regularization

  • evidenced by latent feature similarity (LF-sim) for each pair of users
  • LF-sim tells us how similar a pair of users in the recommender’s viewpoint
  • for each pair of friends, plot its LF-sim before and after social regularization
  • after SR, friends become highly similar
    • linear regression line is far from diagonal

Fairness Concerns

  • social regularization does not impact all users equally
  • if we look at cold and warm users
    • based on observed level of feedback
  • we find that the impact is
    • very strong for cold-cold friends,
    • strong for cold-warm friends
    • moderate for warm-warm friends

How to address this?

  • improve fairness by requiring social regularization only for warm-warm friends
    • no significant impact on accuracy

Diversity Concerns

  • social regularization may lead to social echo chambers
    • a user becomes more strongly influenced from its community

How to measure this?

  • community influence at k (CI@k) of a user is the proportion of their k most similar users that belong to the same community
    • when CI is high, there is little influence coming outside the community
  • the change in CI@k (ΔCI@k) after vs. before SR measures how strong the effect of SR is

How to address this?

  • improve diversity by allowing a user to be dissimilar to the average community member
  • the change in community influence (ΔCI@k) is smaller
    • no significant impact on accuracy