ניתוח נתונים ולמידת מכונה

אמית רננים – חלופה ליחידה 3

מוטי בזר

mottibz@gmail.com

https://www.commridge.com

Linear Regression using

Scikit-Learn

Some of the materials are with permission from

Jose Portilla, Ariel Bar Itzhak, Koby Mike

The Philosophy Behind Scikit-Learn (Sklearn)

• One-stop-shop library containing many machine learning algorithms
• Utilizes a generalized “estimator API” framework for calling models
• Import, fit, and use in a uniform way across all algorithms
• Easy algorithms swap for testing
• Hides algorithms complexity from the users
• Includes many tools (train-test split, cross validation, metrics, etc.)

​

​

2

Supervised Machine Learning (ML) Process (1 of 3)

​

• Reminder on the ML process:

3

Supervised Machine Learning (ML) Process (2 of 3)

4

Train-Test Split

The ML Models building process

(initial steps)

Supervised Machine Learning (ML) Process (3 of 3)

5

The ML Models building process up to deployment

Scikit-Learn generalized “estimator API” framework

6

1. Train-test split

2. Import the desired algorithm

3. Define the model object

4. Perform the fit() operation

5. Perform the predict() operation

​

​

6. Import desired error evaluation function

7. Evaluate model’s performance

This framework will be similar for all Supervised

machine learning based Scikit-Learn algorithms

Exploring multiple features

• In the introduction section, we explored the relationship between “total spend” and “sales”
• Let’s expand that in understanding the relationship between each of the advertising channels and sales
• Let’s go to the notebook (first section)

7

Regression Performance Evaluation Metrics

• Now that we know how to fit and predict, we need a way to understand how good are the predictions
• We will use the test set for that, to compare our predictions to the actual values
• The most common evaluation metrics for regression are:
• MAE: Mean Absolute Error
• MSE: Mean Squared Error
• RMSE: Root Mean Square Error

​

8

Regression Performance Evaluation Metrics

• MAE: Mean Square Error
• Calculate the mean of the absolute values of the errors:

​

​

• But MAE does not punish large errors:

​

• We want the error metric to account �for high errors and punish them

​

• This makes MAE non popular

​

9

Regression Performance Evaluation Metrics

• MSE: Mean Squared Error
• Punishes large errors by squaring the error values:
• The issue with MSE is that it provides different units�than y (it provides y squared). RMSE addresses that

​

• RMSE: Root Mean Square Error
• It calculates the root of the MSE:
• This is the most popular metric as�it provides the same units as y
• When discussing what is a good RMSE, it all depends on the context
• RMSE of $10 is great when predicting a house price, but horrible when predicting the price of a candy

​

10

Regression Performance Evaluation Metrics

• R² is an additional very popular metric for evaluating regression predictions results

11

Sum of all differences between each y and the mean of all y values

Green horizonal line represents the mean of all y values

This is MSE

• Values are between 0-1
• 1 = Best result (SSerror is very small)
• 0 = Worst result (SSerror is very high)
• A negative result means that the prediction is worse compared to the mean

Regression Performance Evaluation Metrics

• Let’s go back to the notebook to calculate the prediction metrics

12