Relative Variable Selection for Neural Network Parameter Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional stock price prediction methods using neural networks are limited by the lack of comprehensive network architectures and parameter selection mechanisms, leading to inaccurate predictions due to too many or too few parameters in the hidden layers.
Innovation Solution
A relative variable selection system comprising a receiving module, selection modules, and calculating modules that sequentially select variables based on correlation coefficients and weighted values to optimize parameter selection, using a stepwise regression correlation selection method and the bees algorithm for neural network training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If too many parameters are used in the hidden layer of a complicated model, then the model can capture more complex patterns, but the network will lack the ability of mathematical induction and become overfitting
Solution Approach 1:
The patent extracts and removes redundant parameters from the neural network model through a systematic selection process. It uses correlation analysis to identify and eliminate variables that do not contribute significantly to the prediction accuracy, thereby reducing overfitting while maintaining the network's ability to capture important patterns.
Solution Approach 2:
The patent implements a dynamic parameter selection mechanism that adapts the number and type of parameters based on the specific characteristics of the input data. The system dynamically determines the optimal hidden layer configuration by evaluating correlation coefficients and information gain metrics, allowing the model to balance complexity and generalization ability according to the actual data distribution.
2Device complexity
If too few parameters are used in the hidden layer, then the network can maintain simplicity and avoid overfitting, but the network will be unable to obtain an accurate prediction result
Solution Approach 1:
The patent performs preliminary analysis of the input variables before constructing the neural network model. It calculates correlation coefficients and information gain metrics in advance to identify the most relevant parameters, ensuring that the selected variables have the highest potential contribution to prediction accuracy before the model training begins.
Solution Approach 2:
The patent systematically varies the number and combination of parameters in the hidden layer to find the optimal configuration. By changing parameters such as the number of hidden neurons, activation functions, and learning rates, the system identifies the specific parameter settings that achieve the best balance between model simplicity and prediction accuracy for the given dataset.
3Measurement precision
If comprehensive network architectures and parameter selection mechanisms are implemented, then prediction accuracy can be improved, but the system complexity and computational burden increase
Solution Approach 1:
The patent divides the parameter selection process into multiple independent stages: initial variable selection based on correlation analysis, feature engineering to create derived variables, and iterative optimization of network architecture parameters. This segmentation allows each stage to be optimized independently, reducing the overall system complexity while achieving comprehensive parameter selection.
Solution Approach 2:
The patent implements an automated parameter selection mechanism that uses the input data itself to determine the optimal model configuration. The system automatically calculates correlation metrics, evaluates information gain, and selects parameters based on the inherent characteristics of the data, eliminating the need for manual intervention and reducing system complexity.
Data Source
AI summary
The present invention discloses a relative variable selection system and a selection method thereof. In the present invention, the receiving module receives a plurality of variables. Based on a correlation coefficient of variables, a first selection module sequentially selects variables with a correlation coefficient greater than a first threshold value. Based on the variables selected by the first selection module, a first calculating module selects a regression value and a weighted value corresponding to the foregoing variables. Based on the weighted values, a second selection module sequentially selects variables with a weighted value smaller than a second threshold value. Based on the variables selected by the second selection module, a second calculating module calculates analyzed values of the foregoing variables. Based on the analyzed values of the variables, a third selection module selects analyzed values which are greater than the target value.