Classification

application

    classification :
        image classification
        identify fraud detection
        customer retention
        diagnostics

topics

• binary ; decision boundary ; accuracy ; hamming loss ;

• linear ; linear boundary - the role of w and b ; reparametrization and approximation using sigmod to define loss function ; gradien and accuracy to perform gradient descent ; validate using complex trick;

• svm ; maximize the margin ; define gap problem ; hard and soft margins(C - hyperparameter) ; train svm to show the effect of C ; show score and accuracy ; optimize C param;

• knn ; k=1 euclidean distance (best overfit) ; cosine dist ; shortest path ; the course of dimensionality

• full training process ; mnist example ; confusion matrix ; hyper params under/over/fit ; how to divide the data ; cross validation - K fold ;

• resample

Performance Scores

• decision function multi class performance evaluation ; one vs all ; one vs one ; (scikit-func)

• recall ; precision ; f1 ; (scikit-func)

• Balanced Accuracy ;

• roc ; auc ; (scikit-func)

  • understand auc lab

• cost ; loss ;

• scikit-classificationReport

• scikit classification metrics

labs

• lab : mnist knn

• lab : mnist svm

• lab :score demo ; Classification Report

• lab :decFunc ; decision function ; predict_proba

• lab : perf score

• lab : cal imbalanced