Dynamic Ensemble Prediction via Selective Model Combination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing ensemble prediction methods face challenges in achieving accurate results due to high average prediction errors caused by large errors in base model predictions, especially when weight calculation accuracy decreases, leading to increased errors in selection methods and higher errors in weighted sum methods.
Innovation Solution
A method and apparatus for selective ensemble prediction based on dynamic model combination, which involves determining an optimal model combination parameter using verification data, calculating model weights for input data, and applying normalization to select and weight model predictions, excluding models with high errors and using Sparse-max to optimize weights for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the best selection method is used to choose predictions from base models, then prediction accuracy is improved when model weights are accurate, but average prediction error increases when weight calculation accuracy decreases
Solution Approach 1:
The patent applies dynamics by transitioning from static weight assignment to dynamic weight adjustment. The system continuously monitors prediction errors and adapts model weights in real-time, allowing the ensemble to respond to changing conditions. This dynamic approach enables the system to maintain high prediction accuracy while reducing average prediction errors by adjusting weights based on current performance rather than relying on fixed or pre-calculated weights.
Solution Approach 2:
The patent implements feedback mechanisms where prediction errors are continuously measured and fed back into the weight calculation process. The system uses the actual prediction performance to adjust model weights, creating a closed-loop control system. This feedback enables the ensemble to learn from mistakes and improve over time, resolving the contradiction between maintaining accuracy and reducing average errors through iterative optimization.
2Reliability
If the weighted sum method is used to combine base model predictions, then errors from individual models are offset, but the average error remains higher than selection methods
Solution Approach 1:
The patent applies parameter changes by dynamically modifying the weights assigned to different base models based on their current performance metrics. Instead of using fixed weights or simple averaging, the system adjusts weight parameters in response to changing conditions and prediction accuracy. This allows the ensemble to leverage the error-offsetting benefit of weighted sums while maintaining lower average errors through adaptive parameter optimization.
3Reliability
If more base models are included in the ensemble, then prediction robustness is improved, but computational complexity and processing time increase
Solution Approach 1:
The patent applies the extraction principle by selectively removing or de-emphasizing underperforming base models from the ensemble based on their prediction accuracy. The system continuously evaluates each model's contribution and extracts only the high-performing models into the active ensemble. This reduces computational complexity while maintaining robustness by focusing resources on the most reliable predictors rather than including all available models.
Solution Approach 2:
The patent implements discarding and recovering by temporarily excluding underperforming models from the ensemble while preserving their ability to contribute when conditions change. The system monitors model performance and can discard poor performers to reduce complexity, then recover them when they demonstrate improved accuracy. This dynamic inclusion/exclusion strategy maintains robustness while managing computational resources efficiently.
Data Source
AI summary
Disclosed are a method and apparatus for selective ensemble prediction based on dynamic model combination. The method of ensemble prediction according to an embodiment of the present disclosure includes: collecting prediction values for input data of each of the prediction models; calculating a model weight of each of the prediction models using a pre-trained ensemble model that uses the prediction value as an input; selecting at least some model weights from the model weights using a predetermined optimal model combination parameter; and calculating an ensemble prediction value for the input data based on the selected model weight and a prediction value of a prediction model corresponding to the selected model weight.