Temporal Knowledge Graph Forecasting via Pattern Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing forecasting methods for Temporal Knowledge Graphs (TKGs) fail to effectively incorporate the structural and relational information, leading to inaccurate predictions of future scenarios due to information overload and limited consideration of graph structure and temporal patterns.
Innovation Solution
A two-step approach for Temporal Knowledge Graph forecasting, involving relation prediction and entity prediction, where relation prediction models learn patterns in relations across time and entity prediction models capture patterns in entities, using sequential data and graph embeddings to forecast future relations and entities, thereby enhancing prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If existing ML-based forecasting methods are used, then computational processing is simplified, but they cannot take into account the relations between entities and the graph structure
Solution Approach 1:
The patent segments the forecasting task into two independent prediction steps: first predicting future relations based on historical relation patterns, then predicting future entities based on predicted relations. This segmentation allows each step to focus on specific aspects of the data, improving both computational efficiency and information utilization.
Solution Approach 2:
The patent transforms the graph structure into a sequential data format by ordering relations and entities across time steps. This dimensional transformation enables the use of sequential models that can capture temporal patterns while maintaining graph structure information, bridging the gap between simplified computation and comprehensive information processing.
2Loss of information
If existing Knowledge Graph-based forecasting methods are used, then relational data and graph structure are considered, but not all available information in TKGs is considered
Solution Approach 1:
The patent performs preliminary action by first predicting future relations before predicting future entities. This sequential approach ensures that relation information is captured and processed before entity prediction, allowing the system to leverage both relational and graph structure information systematically to improve prediction accuracy.
Solution Approach 2:
The patent maintains continuity of useful action by processing information in a continuous sequential manner across time steps. The model continuously updates its predictions by incorporating new temporal information while maintaining the graph structure, ensuring all available information is considered without interruption.
3Ease of operation
If humans process TKG data, then intuitive understanding is achieved, but information overload prevents reliable prediction
Solution Approach 1:
The patent introduces an intermediary computational model that processes the complex TKG data. This intermediary system translates the intuitive but overwhelming human-readable graph data into structured sequential representations that can be processed systematically, achieving both intuitive understanding and reliable prediction by bridging human cognition and computational processing.
Data Source
AI summary
A predictive policing system includes a database of crime related scenarios in a number of past timesteps represented as temporal knowledge graphs (TKGs). Crime prediction devices generate relationship vectors that describe relations for each node of the TKGs for each available timestep. A dataset is created including vector sequences sets for each node of the TKGs, which are used as sequential inputs for training a pattern model to predict future relations for each node of the TKGs. A forecasting model is trained to predict nodes of the TKGs associated with each of the predicted future relations. Predicted future TKGs are assembled describing a crime scenario in an area of interest per future time steps of interest. A forecasting-based action recommendation system computes actions to steer the predicted scenario towards a desired scenario. Monitoring and/or surveillance devices deployed in the area of interest are adapted based on the computed actions.