1. Introduction
A major problem with pattern recognition is the dimension disaster. In chapter 1, we will see that the dimension can easily become very large.
There are several reasons for the necessity to reduce the dimension. Computing complexity is one aspect. The generalization performance of another classifier.
Therefore, the main task of this chapter is to select important and less-dimension features while retaining the classification and discriminant information of features as much as possible. This process is called feature selection or feature dimensionality reduction.
In quantitative description, the selected feature should reduce the intra-class distance and the inter-class distance.
In some documents, feature extraction is used instead of featureselection. This will conflict with some descriptions in Chapter 7th.
2. Preprocessing
① Remove a group value (outlier removal)
The deviation value is defined as a point with a large mean deviation from the relevant variable.
② Data normalization
Many feature values are located in different dynamic thresholds, and the large feature values have a great impact on the loss function. Therefore, normalization to the similarity threshold will help.
③ Missing Data
Some features are missing from the feature quantity. For example, remote sensing is covered by other sensors in a specific area.
3. Peak Value
4. Feature Selection Based on statistical assumptions
5. The receiver operating characteristic curve
6. Classifier
7. feature subset selection
8. Optimal feature generation)
9. Neural Networks and Feature Selection
10. One hint generalization Theory
11. Bayesian information standards
[pattern recognition]. (Greece) ritis Note 5 _ feature selection