ML Prediction Accuracy via Knowledge Graph Feature Injection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-learning models often face reduced prediction accuracy due to insufficient input features, leading to inefficient processing and less accurate results.
Innovation Solution
The implementation leverages knowledge graphs by encoding input data characteristics, including gradients of multiple features, and injecting these encodings into a textual knowledge graph to obtain supplemental data from a higher-feature-count ML model, thereby enhancing prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If fewer features are used in the input data, then processing speed is improved, but prediction accuracy deteriorates
Solution Approach 1:
The system performs preliminary action by proactively querying for supplemental features from knowledge graphs before making a final prediction when the initial confidence level is insufficient. This allows the system to obtain additional relevant features on-demand rather than processing all possible features upfront, thus maintaining fast processing speed while improving prediction accuracy when needed.
Solution Approach 2:
The system dynamically adjusts the number of features used in processing based on the confidence level of the initial prediction. When confidence is sufficient, only the minimal feature set is processed quickly. When confidence is insufficient, the system dynamically queries for and incorporates supplemental features from knowledge graphs, making the feature processing adaptive rather than static.
2Measurement precision
If more features are accounted for in the input data, then prediction accuracy is improved, but processing efficiency deteriorates
Solution Approach 1:
The system applies partial action by using only the necessary number of features required to achieve sufficient prediction confidence. Instead of processing all available features uniformly, the system queries knowledge graphs selectively for supplemental features only when the initial prediction confidence is below the threshold, thus avoiding unnecessary processing overhead while ensuring adequate accuracy.
Solution Approach 2:
The knowledge graph serves as an intermediary that bridges the gap between limited input features and comprehensive prediction accuracy. Rather than directly processing all possible features through the ML model, the system uses the knowledge graph as a mediator to retrieve relevant supplemental features on-demand, improving accuracy without proportionally increasing processing burden.
3Measurement precision
If a higher number of features are used in the second ML model, then supplemental prediction accuracy is improved, but model complexity increases
Solution Approach 1:
The system segments the prediction task into two stages: a first ML model with fewer features for initial prediction, and a second ML model with more features for supplemental prediction only when needed. This segmentation allows each model to be optimized for its specific purpose, reducing the overall complexity burden while maintaining high accuracy when supplemental prediction is required.
Solution Approach 2:
The system employs a nested structure where the second ML model with higher feature count is effectively nested within the overall prediction system, activated only when the first model's confidence is insufficient. This nested arrangement allows the more complex model to be contained and utilized selectively, improving accuracy without permanently increasing the operational complexity of the system.
Data Source
AI summary
Implementations include receiving a predicted value and confidence level from a first ML model, and determining that the confidence level is below a threshold, and in response: providing an encoding based on input data and non-textual information to the first ML model, the encoding representing characteristics of the input data relative to the predicted value, the characteristics including respective gradients of features of the input data, injecting the encoding into a textual knowledge graph that corresponds to a domain of the first ML model to provide an encoded knowledge graph, receiving supplemental data based on the encoded knowledge graph, and providing a supplemental predicted value from a second ML model based on the input data and the supplemental data, the second ML model having a higher number of features than the first ML model, and the supplemental predicted value having a supplemental confidence level that exceeds the threshold.