Temporal Knowledge Graph Forecasting With Explainable Path Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning-based methods for temporal knowledge graph (TKG) forecasting struggle with scalability, lack of inductive approaches, and lack of explainability, particularly in semi-inductive settings, failing to account for temporal and relational paths and distances between nodes.
Innovation Solution
A method that computes spatial-temporal and relational paths to anchor nodes in a TKG, using a vocabulary encoder to determine embeddings based on these paths, and predicts future interactions by scoring these embeddings, providing explainability and scalability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If classic machine learning-based forecasting methods are used for temporal knowledge graphs, then the methods can handle simple time-series data, but they fail to exploit structural dependencies between entities resulting in inaccuracies and poor performance
Solution Approach 1:
The patent introduces path-based representations as intermediaries that connect nodes in the knowledge graph through sequences of relations. These paths serve as mediators that capture structural dependencies between entities, allowing the model to exploit graph structure while maintaining compatibility with machine learning forecasting approaches. The path embeddings act as a bridge between the graph structure and the prediction mechanism.
Solution Approach 2:
The patent segments the knowledge graph into individual paths between node pairs, rather than treating the entire graph as a single complex structure. By decomposing the graph into manageable path segments and computing embeddings for each path independently, the model can process structural dependencies in a computationally efficient manner while maintaining prediction accuracy.
2Productivity
If algorithms are used to predict how large knowledge graphs will change over time, then predictions can be made, but the systems lack inductive approaches and struggle with scalability to new unseen nodes
Solution Approach 1:
The patent transitions from node-centric representations to path-centric representations by introducing an additional dimensional perspective. Instead of representing nodes individually, the model represents relationships between nodes through paths, allowing it to generalize to new nodes by leveraging path patterns observed in the training data. This dimensional shift enables inductive reasoning about unseen entities.
Solution Approach 2:
The patent changes the fundamental parameters of representation from fixed node embeddings to flexible path-based embeddings that can be composed from relation and entity components. This parameter transformation allows the model to generalize to new nodes by combining known relation patterns with new entity identifiers, providing inductive capability without requiring retraining.
3Measurement precision
If path-based methods are used to compute embeddings for all nodes, then comprehensive coverage is achieved, but computational complexity and time requirements increase significantly
Solution Approach 1:
The patent extracts and pre-computes path embeddings for the training set, separating the computationally intensive path extraction task from the prediction task. By pre-computing path representations offline and storing them for reuse, the model eliminates redundant computation during prediction while maintaining comprehensive path-based coverage for accurate embeddings.
Solution Approach 2:
The patent performs preliminary computation of path embeddings and relation encodings before the actual prediction task. By preparing path representations in advance and caching them, the model reduces computation time during inference while maintaining comprehensive path-based analysis for accurate embeddings.
4Loss of information
If detailed temporal and relational path information is computed for each node, then explainability is improved, but device complexity and computational resources increase
Solution Approach 1:
The patent introduces path embeddings as intermediary representations that capture temporal and relational information in a structured, interpretable format. These path embeddings serve as mediators between the complex graph structure and the prediction output, providing explainability by showing which paths and relations contribute to predictions without requiring the full complexity of the original graph structure to be maintained in memory.
Data Source
AI summary
A computer-implemented method for predicting links in a temporal knowledge graph (TKG) includes determining one or more anchor nodes and computing, from each node to each anchor node of the TKG for each time-step, relational and temporal paths, and temporal and spatial distances. An embedding is determined for each node to a closest anchor node at each time-step using a vocabulary encoder that combines information received from separate encoders configured to encode the paths and distances. The embedding includes a type of relation. Scores are predicted for each embedding at a future time-step using a scoring function. Link prediction is performed to predict how interaction of the nodes change at the future time-step based on the scores. The present disclosure has applications including, but not limited to, use cases in computational biology, medical AI and healthcare, and cyber threat security for optimizing machine learning processes or supporting decision making.