guide.txt

when get data  DO THE FOLLOWING THINGS ::


check missing  values 
check duplicate values 
data look mathematically {describe fn } ::  helps to detect outliers , only works for numerical col 
correlation b/w cols ::  helps to remove unrelated col


FINAL TEXT :
----------------------------------------------------------------------------------------
r2 score not applied in classification tasks but used for regression tasks
low bias +  high variance models 
----------------------------------------------------------------------------------------


General Understanding

1) To reduce overfitting :
- bagging - boosting - regularization 

2)feature extraction :
- PCA - LDA(Linear Discriminant Analysis) - T-sne




Charts :  


Univariate : 
countplot 
pieChart 
histogram 
distplot 
boxplot 
bubble charts


Bivariate : 

scatterplot 
BarPlot 
Line plot 
pair plot 
HeatMap 
Cluster Map
JointPlot

Multivariate:

3d scatter plot 
facetgrids 
Pair plots 



=> REGRESSION  : 
Simple Lr :  DIRECT FORMULA / Grdaient descent 
multiple LR 
Polynomial 



=> UNSUPERVISED LEARNING   : 
K means clustering 
agglomerative clustering
DBSCAN 

                                        FEATURE_ENGINEERING
    
     FEATURE SCALING :

          standardization : [ StandardScaler  ]
          Normalization : [ MinMaxSCaler,Mean Normalization , MaxAbsScaling  , RobustScaler]  
     Encoding data : 
          Label Encoder 
          Ordinal Encoder 
          One hot encoding 
          pandas dummies
     workflow transformations  : 
          ColumnTransformer 
          Pipeline 

     Hyperparamter tuning : 
          Cross Val Score || K-fold
          Grid Search CV 
     Missing Values  :
          drop them 
          Simple Imputer 
          KNN imputer 
          Missing Indicator
          Iterative Imputer 
          Numerical COLS : [by mean, by median, random,end of distribution] 
          Categorical COLS : [mode :most frequent , new category : missing]
   
     



Gradient Descent -  batch , Stochastic , mini-batch GD 




PCA 🟢
LINEAR REGRESSION 🟢
-REGULARIZARION:  🟢
|| RIDGE 
|| LASSO 
|| ELASTIC NET 
GRADIENT DESCENT: 🟢
|| MINI BATCH 
|| SGD
|| BATCH 
LOGISTIC  🟢
DECISION TREES 🟢
VOTING ENSEMBLE  🟢
BAGGING  🟢
RANDOM FOREST  🟢
-BOOSTING:
|| ADABOOST 🟢
|| GRADIENT BOOSTING
|| XGBOOST 
STACKING  🟢
KNEIGHBORS 🟢
SVM 
K-MEANS  🟢
AGGLOMERATIVE 🟢
DBSCAN 🟢
NAIVE BAYES(+GNB) 🟢
SMOTE 🟢
OPTUNA 🟢
ROC curve 🟢






SVM :
- constraint optimization {convex analysis and  optimization theory || lagranges's multipliers || quadratic programming}
-  why ||w|| == ||w||^2/2=>  just to make it differentiable 
-  1/||w||  comes from  

stats , vo anova and chi square chod diya hain


Ml : GRADIENT BOOSTING and XGBOOST {superb algo just baad me } {much to explore and experiement} {only campusx videos}{ref}
: SVM {learn constraint optimization theory & lagranges multipliers } + {saved articles and videos}








some other techniques : 

=>  df.isnull().mean() * 100 => gives the percentage of missing data directly 
=> df.dropna().sample() => gives a sample from dataset after dropping null values 
=> np.mean(cross_val_score(mode,x,y,cv,scoring="accuracy"))  => gives the final acc by fitting to the model "cv" times
=> data[col].skew() => if ~0  then data is normally dist else >0  right skewed <0 left skewed , right skewed => outliers on right side