Conceitos essenciais
Explaining neural model predictions creatively for user adoption in enterprise applications.
Resumo
In the NeurIPS 2020 Competition and Demonstration Track, the focus is on xLP: Explainable Link Prediction for Master Data Management. The authors emphasize the importance of providing explanations for neural model predictions to users, particularly in enterprise settings where trust and user adoption are crucial. They highlight the challenges of using Graph Neural Networks (GNN) models in enterprise applications, especially when dealing with sensitive data like customer relationships. The article discusses different explainability solutions drawing from research in interpretability, fact verification, path ranking, neuro-symbolic reasoning, and self-explaining AI to enhance user understanding and comfort with model predictions.
The authors compare three explainability techniques for Link Prediction: interpretable models approximating neural predictions, link verification using external information, and path ranking algorithms previously used for error detection. They explore how users' preferences towards different explanation types can impact their ability to understand models effectively. The work focuses on explaining link prediction in master data management tasks involving entity matching, non-obvious relation extraction, and more within property graphs.
Furthermore, the article delves into related works on graph neural networks, entity matching, and knowledge graph embeddings to provide a comprehensive background. It evaluates the performance of different models on various datasets and emphasizes the need for deploying such systems under professional oversight to ensure ethical considerations are met.
The authors present three human-understandable solutions – search-based retrieval of verification text, anchors-based explanation, and path ranking-based explanations – to elucidate links predicted by Graph Neural Networks. A case study evaluating annotator agreement with these explanations showcases the effectiveness of each technique in enhancing user comprehension.
Estatísticas
UDBMS Dataset ROC AUC: GCN - 0.4689; P-GNN - 0.6456
MDM Dataset ROC AUC: GCN - 0.4047; P-GNN - 0.6473
Citações
"Explaining neural model predictions to users requires creativity."
"While some might be interested in model interpretability, others might want a quick and easy-to-understand solution."
"Our goal here is to understand user preferences towards different types of explanations."
"Link Prediction on people graphs presents a unique set of challenges."
"We want the Data Stewards to draw valuable insights into the model’s learning process."