updated

Author: peicheng

FeatureEngineering.md

@@ -8,28 +8,26 @@ We can combine tables, transform features to create new features.
 We can even do so with help. Let's see [featuretools](https://www.featuretools.com/)
 
 After we created "meaningful" features, we are now ready to transform their format.
-### OneHot Encoding: Dealing with discrete features
 
-### Mathematical Operations: Creating New Features
 
-#### Single Column
-Apply functions such as log, sqrt, pow, or other functions that take 1 input.
-
-#### Two Columns
-Apply product,ratio, or other transformations that take 2 or more inputs.
-
-### Bucketing: Dealing with continuous feature
+## Dimension Reduction (Yin-side)
 
-### NLP: dealing with text feature
+There are many available tools for this.
 
-## Dimension Reduction (Yin-side)
+### Statistics: remove highly correlated data
+We can do this automatically using [featuretools](https://www.featuretools.com/).
+Or we can remove them by hand.
 
 ### Principal Component Analysis (PCA)
+This is a classic and fast method, but it has it's limitations.
+Remember standardize your data before you do this.
+We can do this using sklearn very fast.
 
 ### auto-encoding using deep neural networks
+We need sufficient data to do this well, more complicated than PCA>
 
-### Feature selection
 
+### Feature selection
 Before the final task, we could try to solve a representive task. Use feature importance for a model, usually trees, to select features. Use SelectKBest (e.g. [here](https://scikit-learn.org/stable/modules/generated/sklearn.feature_selection.SelectKBest.html)).
 In addition, we can do selection while training - LASSO regularization etc.