MINDWALC: mining interpretable, discriminative walks for classification of nodes in a knowledge graph

Area of Science:

• Graph Machine Learning
• Knowledge Representation and Reasoning
• Data Mining

Background:

• Machine learning on graphs enhances expressiveness by encoding entity relations.
• Knowledge graphs are multi-relational directed graphs representing domain knowledge.
• Deep learning on knowledge graphs offers high predictive performance but lacks interpretability, crucial for critical domains like healthcare.

Purpose of the Study:

• To present a technique for mining interpretable walks from knowledge graphs for node classification.
• To develop efficient data structures and mining algorithms for these walks.
• To combine mining with multiple classification approaches to balance explainability, performance, and runtime.

Main Methods:

• Mining of specific, informative walks from knowledge graphs.
• Development of efficient data structures for walk mining.
• Integration of walk mining with three distinct node classification strategies.

Main Results:

• The proposed technique was compared against state-of-the-art black-box methods on four benchmark datasets.
• The interpretable approaches demonstrated competitive or superior performance to black-box alternatives.
• Key finding: interpretability was achieved without sacrificing predictive performance.

Conclusions:

• Mining interpretable walks is a viable alternative for knowledge graph node classification.
• This method provides inherently transparent models.
• Achieves strong predictive performance, unlike current deep learning techniques.