Adaptive pseudo-Siamese policy network for temporal knowledge prediction.

Area of Science:

• Artificial Intelligence
• Data Science
• Computer Science

Background:

• Temporal knowledge prediction is vital for early event warning systems.
• Existing methods struggle with modeling fact evolution and handling unseen entities.
• Temporal knowledge graphs are used to predict future facts from historical data.

Purpose of the Study:

• To propose a novel adaptive pseudo-Siamese policy network for temporal knowledge prediction.
• To address challenges in modeling evolutionary patterns and unifying seen/unseen entities.
• To enhance the accuracy of predicting future facts in temporal knowledge graphs.

Main Methods:

• A reinforcement learning-based adaptive pseudo-Siamese policy network is developed.
• Two sub-policy networks are employed: one for entity-relation paths (static patterns) and another for relation-time paths (unseen entities).
• A temporal relation encoder and a gating mechanism are utilized to capture temporal patterns and integrate sub-network results.

Main Results:

• The proposed model demonstrates considerable performance on link prediction tasks across four benchmark datasets.
• Extensive experimental results validate the effectiveness of the adaptive pseudo-Siamese network.
• The method shows significant improvements compared to existing temporal knowledge prediction techniques.

Conclusions:

• The novel adaptive pseudo-Siamese policy network effectively predicts temporal knowledge, outperforming existing methods.
• The approach successfully models evolutionary patterns and handles both seen and unseen entities.
• This work advances early event warning systems through improved temporal knowledge prediction.