Day #82 - Validation and Overfitting

• Train Data (Past), Unseen Test Data (Future)
• Divide into three parts - Train (Past), Validation (Past), Test (Future)
• Underfitting (High Error on Both Training and Validation)
• Overfitting (Doesn't generalize to test data, Low Error on Train, High Error on Validation)
• Ideas (Lowest Error on both Training and Testing Data)

Validation Strategies

• Hold Out (divide data into training / testing, No overlap between training / testing data ) - Used on Shuffle Data
• K-Fold (Repeated hold out because we split our data) - Good Choice for medium amount of data, K- 1 training, one subset - Used on Shuffle Data
• Leave one out : ngroups = len(train) - Too Little data (Special case of K fold, K = number of samples)
• Stratification - Similar target distribution over different folds

Stratification useful for

• Small datasets (Do Random Splits)
• Unbalanced datasets
• Multiclass classification

 Stratification preserves the target distribution over different folds

	#Tip #1
	sklearn.model_selection.ShuffleSplit
	#Tip #2
	sklearn.model_selection.Kfold
	#Tip #3
	sklearn.model_selection.LeaveOneOut

view raw Datavalidation.py hosted with ❤ by GitHub

Happy Coding and Learning!!!