An Inductive Reasoning Model based on Interpretable Logical Rules over temporal knowledge graph.

Area of Science:

• Artificial Intelligence
• Data Science
• Knowledge Representation

Background:

• Temporal Knowledge Graphs (TKGs) are crucial for predicting future events.
• Current methods like embedding-based and logical rule-based approaches have limitations in interpretability and scalability.
• There is a need for advanced models that can effectively extrapolate future events in TKGs.

Purpose of the Study:

• To propose a novel hybrid model, ILR-IR, for enhancing future event prediction in TKGs.
• To combine the strengths of embedding-based and logical rule-based methods for interpretable and scalable TKG extrapolation.
• To improve the accuracy and generalization capabilities of TKG reasoning models.

Main Methods:

• Developed the Inductive Reasoning Model based on Interpretable Logical Rule (ILR-IR), a hybrid approach.
• Integrated deep causal logic by extracting insights from logical rules and entity interaction preferences.
• Incorporated a one-class augmented matching loss for enhanced model training and performance.

Main Results:

• ILR-IR demonstrated superior performance in TKG extrapolation compared to state-of-the-art baselines on ICEWS datasets.
• The model exhibited strong generalization capabilities and robust zero-shot reasoning abilities on related datasets.
• Experimental results validated the effectiveness of the hybrid approach for interpretable TKG reasoning.

Conclusions:

• The proposed ILR-IR model effectively addresses the limitations of existing methods for TKG extrapolation.
• ILR-IR offers a promising direction for interpretable and accurate future event prediction in temporal knowledge graphs.
• The model's generalization and zero-shot capabilities highlight its potential for real-world applications.