Enhancing Recommendation Diversity in Trust-Based GNNs Through Adversarial Regularisation

Although the recommender systems based on social trust may be correct, a major issue with them is that they tend to increase popularity bias and form an echo chamber due to their strong tendency to recommend mainstream content in the first place. In this paper, the audit of a baseline trust-aware Graph Neural Network (GNN) is used to measure this bias and a new framework to reduce it is proposed. We suggest an Adversarial regularisation mechanism, which re-formulates the training process into a competitive game. In this type of system, the GNN recommender will have to race an opponent network, which we define as the Discriminator. This Discriminator is trained to identify the biased patterns of the recommenders, which will force them to come up with more diverse and less predictable recommendations. Our empirical findings confirm the effectiveness of the approach, as there was an absolute variation of twenty percent in Popularity Bias that reduced in the baseline model and reached 39.90% to 19.90%. Importantly, this remarkable gain in fairness was attained without worsening the predictive performance, indicating a feasible direction in the creation of systems that are highly performant yet foster content discovery.