Knowledge Graph Event Sequence Prediction With Dynamic Embedding Updates

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for predicting and classifying event sequences in knowledge graphs are computationally intensive, require substantial re-computation and re-training, and fail to provide real-time updates and account for event order.

Innovation Solution

A three-layered model comprising a graph embedding layer trained unsupervised, an event embedding layer trained end-to-end supervised, and a prediction layer, allowing dynamic updates with new events without full re-training, using a transformer-based sequence model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If graph embedding is re-computed for each new event, then prediction accuracy is maintained, but computational time and processing power increase substantially

Engineering Contradiction:
Improveprediction accuracyVSAvoidcomputational time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The knowledge graph embedding is computed once during initialization and stored. When new events arrive, the system uses the pre-computed embedding rather than re-computing it, thereby maintaining prediction accuracy while avoiding the substantial computational time cost of re-computation.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system creates a copy of the knowledge graph embedding that remains static while new events are processed separately. This allows the prediction model to use the pre-computed embedding as a reference without modifying or re-computing it, thus reducing computational overhead while maintaining accuracy.

Inventive Principle:
Principle #26Copying

2Measurement precision

If the prediction model is re-trained when new events arrive, then prediction accuracy is updated, but processing power and computation time increase

Engineering Contradiction:
Improveprediction accuracyVSAvoidprocessing power
Core Design Contradiction:
Measurement precisionVSPower

Solution Approach 1:

The model is trained in advance on historical data and event sequences before new events arrive. When new events occur, the system uses the pre-trained model for predictions without triggering re-training, thus maintaining accuracy while avoiding the high processing power requirements of re-training.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system dynamically adapts to new events by updating only the prediction outputs using the pre-computed embedding and pre-trained model parameters, rather than re-training the entire model. This allows the system to respond to new data in real-time without the computational burden of full model re-training.

Inventive Principle:
Principle #15Dynamics

3Speed

If real-time updates are implemented when new events occur, then system responsiveness is improved, but computational overhead increases

Engineering Contradiction:
Improvesystem responsivenessVSAvoidcomputational overhead
Core Design Contradiction:
SpeedVSLoss of time

Solution Approach 1:

The knowledge graph embedding is prepared in advance and stored for quick retrieval. When new events arrive, the system can immediately use the pre-computed embedding to generate predictions, achieving real-time responsiveness without the computational overhead of on-demand embedding computation.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system uses a static copy of the knowledge graph embedding that can be quickly accessed and applied to new events without re-computation. This allows real-time updates and predictions while minimizing computational overhead since the embedding does not need to be recalculated for each new event.

Inventive Principle:
Principle #26Copying

4Measurement precision

If event sequence order is considered in predictions, then prediction accuracy is improved, but model complexity increases

Engineering Contradiction:
Improveprediction accuracyVSAvoidmodel complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The model is pre-trained to recognize and process event sequence patterns and temporal relationships during the training phase. Once trained, the model can efficiently handle new event sequences by applying the learned patterns without adding complex computational mechanisms, thus maintaining accuracy while avoiding increased model complexity.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The model uses parameterized functions that can adapt to different event sequence lengths and patterns during inference. By using learnable parameters rather than fixed complex structures, the model can accommodate event order information flexibly without increasing structural complexity, allowing accurate predictions while keeping the model relatively simple.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4315166B1Method and system for predicting and classifying event sequences embedded in a knowledge graph
Publication Date: 2025.11.19 NEC CORP
  • EP4315166B1 patent drawingFigure 1
  • EP4315166B1 patent drawingFigure 2
  • EP4315166B1 patent drawing

AI summary

A computer-implemented method for event sequence forecasting of a process instance is disclosed. The method comprises building up and training a three- layered prediction model (160) including a first, a second and a third layer. The first layer is a graph embedding layer that assigns a fixed-dimensional graph embedding vector to each node and relation type in a fused event and knowledge graph (200) that contains available structural information including events, knowledge graph nodes, and the links between them. The second layer is an event embedding layer that assigns to each event of the process instance a fixed-dimensional event embedding vector. The third layer is a prediction layer that receives as input a sequence of event embeddings from the second layer and that generates as output a prediction of an unknown property of the event sequence used as input.