Related Experiment Videos
Data mining of inputs: analysing magnitude and functional measures
1Department of Information Engineering, School of Computer Science and Engineering, University of New South Wales, Sydney, Australia. tom@cse.unsw.edu.au
International Journal of Neural Systems
|April 1, 1997
Summary
This study introduces new methods for feature selection in neural networks by analyzing weight matrices and functional input contributions. These techniques improve data encoding and identify significant inputs for better model training.
Area of Science:
- Machine Learning
- Artificial Intelligence
- Data Science
Background:
- Data encoding and feature selection are critical challenges in training back-propagation neural networks.
- Real-world data often contains irrelevant or redundant features and imbalanced class distributions.
- Previous methods struggle with pre-processed and noisy datasets.
Purpose of the Study:
- To investigate novel techniques for feature selection in neural networks using weight matrix analysis and functional measures.
- To evaluate the significance of input features by considering their unique contribution to network outputs.
- To compare new methods against existing literature and validate their effectiveness on real-world data.
Main Methods:
- Analysis of the trained neural network's weight matrix to identify significant inputs.
- Introduction and comparison of a new weight matrix analysis technique with existing methods.
- Examination of functional measures to assess input contributions, prioritizing unique information over redundant magnitude.
- Comparison with sensitivity analysis for input significance.
- Application of a novel functional analysis technique to the weight matrix and comparison with training/test data.
- Introduction of a novel aggregation technique.
Main Results:
- The study demonstrates the effectiveness of weight matrix analysis and functional measures for feature selection.
- New techniques were compared against existing methods, showing promising results on satellite and terrain model data for forest supra-type prediction.
- Functional analysis identified inputs with minor but unique information as more significant than those with high but redundant contributions.
Conclusions:
- Weight matrix analysis and functional measures offer robust approaches to feature selection for neural networks.
- The proposed methods effectively handle noisy and complex real-world datasets.
- This research contributes novel techniques for improving data encoding and input significance determination in machine learning models.