IMF: Interpretable Multi-Hop Forecasting on Temporal Knowledge Graphs.

Area of Science:

• Artificial Intelligence
• Data Science

Background:

• Temporal knowledge graphs (KGs) are increasingly important for applications like event prediction.
• Current KG forecasting models lack interpretability in comparing reasoning paths and use separate networks for different hops, leading to inconsistent scoring of identical semantics.
• This inconsistency hinders the ability to equally measure semantics across different reasoning steps.

Purpose of the Study:

• To develop an Interpretable Multi-Hop Reasoning (IMR) framework for temporal KG forecasting.
• To enhance the interpretability of reasoning path comparisons in temporal KGs.
• To achieve state-of-the-art performance in temporal KG forecasting with improved explainability.

Main Methods:

• The IMR framework transforms multi-hop path reasoning into a stepwise question-answering process.
• It introduces three novel indicators: question matching degree, answer completion level, and path confidence.
• The framework is instantiated using common embedding models like TransE, RotatE, and ComplEx.

Main Results:

• IMR provides a uniform method for integrating paths from different hops based on consistent criteria.
• The framework offers interpretable reasoning paths and explains the basis for path comparisons.
• Instantiated models achieve state-of-the-art performance on four baseline datasets while offering superior explainability.

Conclusions:

• The IMR framework significantly improves the interpretability of temporal KG forecasting models.
• By adopting a consistent, stepwise question-answering approach, IMR overcomes the limitations of black-box path comparisons and inconsistent hop evaluations.
• IMR demonstrates that enhanced explainability can be achieved without sacrificing predictive performance, setting a new benchmark in the field.