KNN Classifiers

what is data imbalance?

imbalanced calssification is the problem of classification when there is an unequal distribution of classes in the training set

• Classification problems are sensitive to class imbalance

  • majority class = most observations (may favor)

  • minority class = least amount of observations

Norma;:300 Diabetes: 250

What are the classification metrics listed?

1. Accuracy

2. Precision

3. Recall

4. F1 Score

What is the accuracy classification metric?

Accuracy: Measures the total number of correct predictions

• ACCURACY DOES NOT PENALIZE INCORRECT PREDICTIONS

Accuracy = Correct Predictions (TP+TN) / (TP+FP+TN+FN_

What is the classification metric precision?

Precisions is a measure of the total predicted positive compared to all psotitives

TP / TP + FP

TP / TP + FN (postives / all positives-right/wrong)

What is the classification metric recall?

Recall is a measurement of true positives compared to model predictions

TP / TP + FN (postives / all actual positives)

What is the classification metric F1 Score?

F1 Score- combines Precision and Recall- mean between

2 Precision * Recall / Precisions + Recall

What are the three sampling techniques?

1. Oversampling: Duplicating values in the minority class to match the majority class

   1. increasing size of minority class to match majority

2. Undersampling: reducing the majority class to match minority class

   1. reducing majority class to match minority

3. Synthetic minority oversampling techniques: SMOTE: creates data points in the data similar to points in the minority class

   1. Oversampling but creating data points from the minority class to add back into the minority class

Whats the difference between classification models and MLR models?

What are the classification models?

Classification models predict a class of membership instead of a continous numerical value

1. Binary Classification:  1 class (yes/no)

2. Multi Class: Multiple classes with exclusive memberships 

3. Multi Label: Multiple calsses that do not have mutually exclusive membership

What are the two things that are a necessity to creating a classification model?

1. data features

2. corresponding labels

Predicting a classes based off given features

Overarching idea of K Nearest Neighbors

determining the most likely class based off closest neighbors

What do you use to determine distance from neighbors?

What is the value of K?

K : the number of nearest neighbors to determine the class of membership

What factors have an impact on the class that is determined?

• The class with the closest and the largest number of neighbors will determine the model’s classification

• The majority class is the one with the most observations

• minority class is the one with the fewest observators

features need to be on a continous numerical scale for KNN and Euclidean distance

IN what scenario does KNN perform poorly and what needs to happen?

Data Nornmalization

• dealing with multiple features that use different scale

• all data should be normalized on a scale 0 to1

• rescale does not change the distribution of the data

what needs to be done to test the model?

data partioning splits the data into training and validation sets

• training data is used to create the model

• validation data is used to evaluate the models accuracy

  • SKLearn train_test_split function partioning data

  • 80/20

what are the variables used for data partioning?

Predictors- variables used as predictive values for KNN must be numeric

Cateogorical- must be one hot encoded

Target- dependent variable that is predicted with KNN exclusive class of membership

whats important to remeber about training data

after data partioning only the training data should be used to train the model

• when creating the model, the value of K must be specified

What are the metrics described?

Accuracy: number of correct predictions

Precisions: number of positives that actually belong to positives (True / True + False Positive)

Recall: number of positives out of all the actual positives (True / True + False Negatives)

F1 Score: mean between precision and recall

What is plain linear regression?

use of current and past data to predict future data

Regression: using a varaible (dependent) and one or more other variables (independent)

Linear Regression: Given a set of observations, determine the equation of line that can be best used to describe dataset

What is the equation for a linear model and the values?

X: independent values (predictor) variables used as predictive values depicted as a value of x (numerical/cateogorical one hot encoded)

Y: dependent values (target)

y = mx + b (slope-m) (intercept-b)

What is exploratory vs predictive modeling?

exploratory: create a model to explain how x and y are related

predictive: can we create a model to predict future values of y and x

• Split observations into training set and validation set

• training-create

• validation-evaluate accuracy

• 80/20 rule

What is the supervised machine learning model?

What are key factors in the validation set?

• validation set is a subset of data randomly sampled from the original dataset

• We use the predictors from the validation set as inputs into a trained model

• model produces estimates of the target

• orginal validation set has known target values(expected values)

• we can compare the target estimates with the known target values to calculate the accuracy or error of the model

What goes into comparing the models accuracy?

the y values from the validation set observations, become the expected values

compare the expected values aganist the values computer from the model

What are the common measurements of error?

Error: absolute of expected - estimated

Mean Error: Average of the errors

Mean Square Error: the sum of the errors squared, divided by the number of errors

Root Mean Square Error: square root of the mean squared

What happens when there is more than one independent variable?

multiple linear regression for independent variables

What does fitting a model mean for multiple linear regression?

solving for the best values of the coefficients

solving for n + 1

can compare the predicted y to the actual y to understand accuracy of the model

How do you select the best predictors for the model?

1. Brute Force Method

   1. Use all the predictors, measure the error

   2. Use a subset of predictors, measure the error

   3. Change the list of predictors, until smallest error

Choose predictors with strongest correlation

What are the steps for using multiple linear regression

1. Import Libraries

   1. Pandas

   2. Scikit-learn

2. Import Data

3. Partition data into training and validation sets

4. Fit training data to a linear model

5. Use the model with validation set

6. Evaluate the model

What are the specific libraries that are needed to be imported

• Scikit Learn: data analytics, data science, machine learning

MLR

• pandas

• train_test_split → sklearn.model_selection

• LinearRegression → sklearn.linear_model

• Mean_Square_error - sklearn.metrics

What are two things important to remeber when importing data?

• data in the columns should be numeric

• Predictors are assumed to be independent of each other

What is the third step detailed of spliting the data into training and validation sets?

1. Create a variable list of predictors. Predictors = []

2. List of predictors, creating a new dataframe of our X values

3. Target =. [list names]

4. Use train_test_split creating training and validation sets

   1. test_size = .2

   2. random_state = 1

How do you fit a model and how do you predict?

df.fit(training x, training y)

df.predictor(list of x values)

How do you evaluate the model?

Use X values from validation set

Compare to the Y values in Validation set

• Mean Error

• Mean Square Error

• Root Mean Square Error

What is R2?

How well the model is fit to the data

• how much variation in target Y can be explained by predictors X

• 1.0 = all variations can be explained (better to 1, the model is better fit for the data)

• dfname.score(validation_x, validation_y)

78% of y can be explained by the predictors used to train the model

What is pandas slicing?

Selecting a subset of data

• Can use ‘loc’ with a boolean mask

age_df = df.loc[df[‘Age]>20