xLightGCN: Explainable Graph Neural Network Based Recommendation Systems with Most Influential Peers

Key points

• Recommendation systems, systems that predict a user's preference for items in order to curate that user's feed, are a ubiquitous and central feature in online social networks.1
• We argue that in the risk-sensitive sociotechnical context of user-facing AI such as recommendation systems, explainability is a key pillar for establishing the trustworthiness of models and validating their effectiveness.1
• We propose a technique built on LightGCN like architectures, an influential and popular class of recommendation models.1
• We describe our approach and argue that it provides an appropriate assurance mechanism for these AI systems leading to specific advantages, including interpretable explanations and a theoretically guaranteed measure of explanation fidelity.1

Sources (1)

• [1]xLightGCN: Explainable Graph Neural Network Based Recommendation Systems with Most Influential Peers
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 02:34 AM

  In this paper we present an exact, graph-based approach for generating explanations for decisions made by recommendation systems.

  Recommendation systems, systems that predict a user's preference for items in order to curate that user's feed, are a ubiquitous and central feature in online social networks.

Extractive summary: sentences quoted from the sources.