Prediction Model Search Space Filtering for Faster Model Construction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing prediction model construction methods consume significant processing resources and are inefficient due to the need to train a large quantity of machine learning models, leading to prolonged determination times.
Innovation Solution
A method that constructs a prediction model by obtaining a targeted model search space based on a dataset and scenario information, filtering models and hyperparameters to reduce the number of selectable options, and training these efficiently to determine the optimal model quickly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all machine learning models in the model search space are trained to ensure comprehensive model evaluation, then the reliability of the prediction model is improved, but the processing resources consumed and the time required increase significantly
Solution Approach 1:
The patent applies preliminary action by performing model filtering before training. Specifically, it calculates feature importance scores for different features in the dataset, uses these scores to filter the model search space to identify a subset of relevant models, and then trains only on this filtered subset. This preliminary filtering action reduces the number of models that need to be trained while ensuring that the most relevant models are evaluated, thus maintaining reliability while reducing time consumption.
2Adaptability or versatility
If the model search space includes all currently collected machine learning models to ensure comprehensive coverage, then the adaptability of the prediction model is improved, but the device complexity and processing resources required increase
Solution Approach 1:
The patent applies local quality by making different parts of the model search space have different treatment based on their relevance. Instead of uniformly processing all models, it calculates feature importance scores and uses these to differentially filter models - keeping models that are relevant to important features while filtering out models that are not. This creates a non-uniform, localized approach to model selection that maintains adaptability for relevant models while reducing complexity by eliminating irrelevant ones.
3Manufacturing precision
If hyperparameter optimization is performed on all machine learning models to ensure optimal performance, then the manufacturing precision of the prediction model is improved, but the processing resources and time consumed increase
Solution Approach 1:
The patent applies the taking out principle by extracting and removing unnecessary hyperparameter optimization steps. Instead of performing hyperparameter optimization on all models in the search space, it first filters models based on feature importance scores, then performs hyperparameter optimization only on the filtered subset of relevant models. This extraction of the optimization process to only necessary models reduces processing resource consumption while maintaining precision for the models that matter most.
Data Source
AI summary
A method, an apparatus, and a computing device for constructing a prediction model, and a storage medium are disclosed, and relate to the field of artificial intelligence technologies. The method includes: obtaining, based on a target dataset of a target prediction scenario and/or scenario information of the target prediction scenario, model search space corresponding to the target prediction scenario; performing model training based on the target dataset and models and hyperparameters that are included in the model search space, to obtain trained prediction models; and obtaining, based on evaluation results of the trained prediction models, a prediction model corresponding to the target prediction scenario. Efficiency of constructing the prediction model can be improved.