辅导 Introduction to Data Science COM SCI X 450.1辅导 留学生R语言

Introduction to Data Science COM SCI X 450.1

CLASS PROJECT

Supervised Machine Learning - Regression

Overview

The class project is designed for you to become productive with your new data science skills by working though a real-life problem using a California Housing data set. We’ll use the data set to predict median house values. This data set appeared in a 1997 paper titled Sparse Spatial Autoregressionsby Pace, R. Kelley and Ronald Barry, published in the Statistics and Probability Letters journal. The   researchers built it using the 1990 California census data. It contains one row per census block group. A block group is the smallest geographical unit for which the U.S. Census Bureau publishes sample data (a block group typically has a population of 600 to 3,000 people). The variables found in the data set areas    follows (the names are fairly self-explanatory, and all variables are atomic class numeric with the exception of ocean_proximity which is a factor):

longitude

latitude

housing_median_age

total_rooms

total_bedrooms

population

households

median_income

median_house_value

ocean_proximity

Each row pertains to a group of houses representing medians for groups of houses in close proximity.

The goal for this project is for you to gain experience in trying out the principles of data science using R as discussed in class. You’ll need to carryout each of the tasks outlined below that parallel the Data Science Process detailed in class.

1. Access the Data Set

The first step is to access the data set and load it intothe R environment. Follow these steps in order to complete this step:

a)  Download the housing.csv data set available in Canvas. The data set has 20,640 rows and 10 variables.

b)  Read the data set into R using a data frame named housing. Please do not use RStudio’s data import feature, but rather write R code for accessing the data.

c)  Cast the ocean_proximity character variable to the factor class, and display the resulting levels.

2. EDA and Data Visualization

Now, let’s do some exploratory data analysis (EDA) and visualizations. Follow these steps below and provide commentary for what you find:

a)  Run the head() and tail() functions on the data frame to get a feel for the actual data values.

b)  Run the summary() function on the data frame. to get a sense for the data classes, range of values for numeric variables, and any NAs found.

c)  Perform. a correlation analysis on numeric variables in the data frame.

d)  Create histograms for each numeric variable.

e)  Produce bloxplots for each numeric variable.

f)   Produce boxplots for the variables: housing_median_age, median_income, and median_house_value “with respect” to the factor variable ocean_proximity.

3. Data Transformation

The nextstep is to transform. the raw data into a more refined form as indicated in the steps below (that will constitute your data pipeline):

a)  We see from the summary() results above that there are many NA values  in the total_bedrooms variable (the only variable with missing values).   This needs to be addressed by filling in missing values using imputation. You can use the “statistical median” for missing total_bedrooms values.

The median is used instead of mean because it is less influenced by extreme outliers. This may not be the best method, as these missing values could represent actual buildings (e.g. a warehouse) with no bedrooms, but  imputation often makes the best of a bad situation. You can use the impute()function covered in class, or write code to accomplish the requirement.

b)  Split the ocean_proximity variable into a number of binary categorical variables consisting of 1s and 0s. Although many machine learning algorithms in R can handle categorical data stored in a factor variable, but we will cater to the lowest common denominator and do the splitting ourselves. Once you’redone with the splitting, you can remove the ocean_proximity variable from the data frame.

c)  Use the total_bedrooms and total_rooms variables along with households to create two new variables: mean_bedrooms and mean_rooms as these are likely to be more accurate depictions of the houses in a given group. You can then remove the total_bedrooms and total_rooms variables once you’ve accomplished this requirement.

d)  Perform. feature scaling. Scale each numerical variable except for median_house_value (as this is our response variable), and the binary categorical variables.

e)  The result of your data transformation efforts should yield a new data frame named cleaned_housing with the following variables:

4. Create Training and Test Sets

Now we can get ready for machine learning by creating the training and test sets using a random sample index.

a)  Create a random sample index for the cleaned_housing data frame.

b)  Create a training set named train consisting of 70% of the rows of the cleaned_housing data frame.

c)  Create a test set named test consisting of 30% of the rows of the cleaned_housing data frame.

5. Supervised Machine Learning - Regression

In this step, you’ll use the randomForest() algorithm found in the randomForest package for training and inference. Our goal is to predict the median house value using regression methods.

First, you’ll need to separate your training set train into two pieces: train_x and train_y where train_x is a data frame. that has all variables except the response variable  median_house_value and train_y is a numeric vector

(not a data frame) that has only the response variable values from median_house_value. In class we used the “formula” method for calling machine learning algorithms, but some algorithms (another otherlanguages) require you to pass the data separately: response variable and predictors, so this exercise will be good practice.

Next, you’llcall the randomForest() algorithm, passing to it both components of the training set created above. Make sure you specify arguments ntree=500, and importance=TRUE. Return the resulting model in the object variable rf as in:

rf = randomForest(x=train_x, y=train_y ,

ntree=500, importance=TRUE)

Now use names(rf) to see all the different metrics computed by the algorithm.

6. Evaluating Model Performance

The final step of the project involves determining the quality of the fit for the

statistical model, i.e. evaluating the performance of the random forest algorithm used above.

a)  Calculate the root mean squared error (RMSE) for the trained model. You

can use the object element rf$mse for this purpose. This element is a vector of calculated mean squared error (MSE) values, one for each “tree” used in the algorithm. You can use the last MSE value,i.e. the last element of the vector. Be sure to calculate the square root of this value to obtain the RMSE. The resulting RMSE is your prediction of median price of a house in a given district to within a RMSE delta of the actual median house price.   This can serve as your benchmark moving forward as you experiment with   other statistical models.

b)  Next, we can see how well the model makes predictions by using the test

set. Split the test set in the same manner as the training set above, creating

a new data frame test_x and numeric vector test_y. You can use the  predict() function using the trained model rfalong with test_x in order to calculate a vector of predicted median house values.

c)  The last step is to calculate the RMSE for the test set using the vector of

predicted median house values and the actual values from the test set,

test_y. To calculate RMSE you can use the UDF we described in class, or you can write your own code.

d)  How does the test set RMSE compare with the training set RMSE? Did the  model score roughly the same on the training and testing data, suggesting that it is not overfit and that it makes good predictions?

e)  Run the variable importance plot with the varImpPlot() function for  the random forest model and discuss what it indicated about your feature vector (you may choose to re-train your model using only the feature variables suggested by this plot.

7. Project Report

Write a project report designed to communicate to a non-technical project  stakeholder the results of the project. Please include the following sections:

•   Project overview: goals and hypothesis. Communicate the “business question.”

•   Description of the data set and all variables.

•   Discussion of results for your EDA and data visualization experiments (please include graphic images of all the plots in this section) .

•   Description of your data pipeline: data munging step.

•   Description of the statistical model you used (random forest) including your feature vector and response variable.

•   Description of the performance metric results for the model.

•   Communicate the “business answer” for the project.

Please provide your report in PDF format.

8. Project Submission

Please use Canvas to submit two files for grading purposes: your R script, and the project report.

9. Congratulations!

Great job, you’ve just gone through the entire Data Science Process to create a supervised machine learning model to predict the response variable

median_housing_value. You now have data science “superpowers!”