Common characteristic engineering operations

Common engineering features include ：

exception handling ：

1. Through the box diagram （ or 3-Sigma） Analyze and delete outliers ;
2.BOX-COX transformation （ Dealing with biased distribution ）;
3. Long tail truncation ;

Feature normalization / Standardization ：

1. Standardization （ Convert to standard normal distribution ）;
2. normalization （ Catch and change to [0,1] Section ）;
3. For power-law distribution , transformation

The data is divided into buckets ：

Equal frequency bucket ;
Equidistant barrel ;
Best-KS Points barrels （ Similar to the use of Gini index for two categories ）;
Chi square is divided into barrels ;

Missing value processing ：

Don't deal with （ For similar XGBoost Tree models, etc ）;
Delete （ Too much missing data ）;
Interpolation completion , Including the mean / Median / The number of / Modeling predictions / Multiple imputation / Compressed sensing completion / Matrix complement ;
Separate boxes , Missing value a box ;

Characteristic structure ：

Construct statistical features , Report count 、 Sum up 、 The proportion 、 Standard deviation ;
Time characteristics , Including relative time and absolute time , The holiday season , Weekends, etc ;
Geographic Information , Including sub boxes , Distributed coding and other methods ;
Nonlinear transformation , Include log/ square / Root sign, etc ;
Feature combination , Feature crossover ;
opinion , Wise men see wisdom. .

Feature screening

Filter type （filter）： First, select the features of the data , Then train the learner , Common methods are Relief/ Variance selection / Correlation coefficient method / Chi square test / Mutual information method ;
Parcel type （wrapper）： Directly take the performance of the learner to be used as the evaluation criterion of feature subset , Common methods are LVM（Las Vegas Wrapper）
; The embedded （embedding）： A combination of filter and wrap , The feature selection is carried out automatically in the process of learner training , Common are lasso Return to ;

Dimension reduction

PCA/ LDA/ ICA;
Feature selection is also a dimension reduction