Machine Learning Engineer Nanodegree

Supervised Learning

Project: Building a Student Intervention System

Welcome to the second project of the Machine Learning Engineer Nanodegree! In this notebook, some template code has already been provided for you, and it will be your job to implement the additional functionality necessary to successfully complete this project. Sections that begin with 'Implementation' in the header indicate that the following block of code will require additional functionality which you must provide. Instructions will be provided for each section and the specifics of the implementation are marked in the code block with a 'TODO' statement. Please be sure to read the instructions carefully!

In addition to implementing code, there will be questions that you must answer which relate to the project and your implementation. Each section where you will answer a question is preceded by a 'Question X' header. Carefully read each question and provide thorough answers in the following text boxes that begin with 'Answer:'. Your project submission will be evaluated based on your answers to each of the questions and the implementation you provide.

Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. In addition, Markdown cells can be edited by typically double-clicking the cell to enter edit mode.

Question 1 - Classification vs. Regression

Your goal for this project is to identify students who might need early intervention before they fail to graduate. Which type of supervised learning problem is this, classification or regression? Why?

Answer: This is a textbook classification problem, as we are trying to classify whether or not the students need additional support. Had we been asked to compute the likelihood of students failing, and therefore needing additional support, that would have been a regression problem.

Exploring the Data

Run the code cell below to load necessary Python libraries and load the student data. Note that the last column from this dataset, 'passed', will be our target label (whether the student graduated or didn't graduate). All other columns are features about each student.


In [1]:
# Import libraries
import numpy as np
import pandas as pd
from time import time
from sklearn.metrics import f1_score

# Read student data
student_data = pd.read_csv("student-data.csv")
print "Student data read successfully!"


Student data read successfully!

Implementation: Data Exploration

Let's begin by investigating the dataset to determine how many students we have information on, and learn about the graduation rate among these students. In the code cell below, you will need to compute the following:

  • The total number of students, n_students.
  • The total number of features for each student, n_features.
  • The number of those students who passed, n_passed.
  • The number of those students who failed, n_failed.
  • The graduation rate of the class, grad_rate, in percent (%).

In [2]:
# number of students
n_students = student_data.shape[0]

# number of features (excluding labeled feature)
n_features = student_data.shape[1] - 1

# passing student count
n_passed = student_data.loc[student_data["passed"]=="yes"].shape[0]

# failing student count
n_failed = student_data.loc[student_data["passed"]=="no"].shape[0]

# graduation rate
grad_rate = float(n_passed) / n_students * 100

# Print the results
print "Total number of students: {}".format(n_students)
print "Number of features: {}".format(n_features)
print "Number of students who passed: {}".format(n_passed)
print "Number of students who failed: {}".format(n_failed)
print "Graduation rate of the class: {:.2f}%".format(grad_rate)


Total number of students: 395
Number of features: 30
Number of students who passed: 265
Number of students who failed: 130
Graduation rate of the class: 67.09%

Preparing the Data

In this section, we will prepare the data for modeling, training and testing.

Identify feature and target columns

It is often the case that the data you obtain contains non-numeric features. This can be a problem, as most machine learning algorithms expect numeric data to perform computations with.

Run the code cell below to separate the student data into feature and target columns to see if any features are non-numeric.


In [3]:
# Extract feature columns
feature_cols = list(student_data.columns[:-1])

# Extract target column 'passed'
target_col = student_data.columns[-1] 

# Show the list of columns
print "Feature columns:\n{}".format(feature_cols)
print "\nTarget column: {}".format(target_col)

# Separate the data into feature data and target data (X_all and y_all, respectively)
X_all = student_data[feature_cols]
y_all = student_data[target_col]

# Show the feature information by printing the first five rows
print "\nFeature values:"
print X_all.head()


Feature columns:
['school', 'sex', 'age', 'address', 'famsize', 'Pstatus', 'Medu', 'Fedu', 'Mjob', 'Fjob', 'reason', 'guardian', 'traveltime', 'studytime', 'failures', 'schoolsup', 'famsup', 'paid', 'activities', 'nursery', 'higher', 'internet', 'romantic', 'famrel', 'freetime', 'goout', 'Dalc', 'Walc', 'health', 'absences']

Target column: passed

Feature values:
  school sex  age address famsize Pstatus  Medu  Fedu     Mjob      Fjob  \
0     GP   F   18       U     GT3       A     4     4  at_home   teacher   
1     GP   F   17       U     GT3       T     1     1  at_home     other   
2     GP   F   15       U     LE3       T     1     1  at_home     other   
3     GP   F   15       U     GT3       T     4     2   health  services   
4     GP   F   16       U     GT3       T     3     3    other     other   

    ...    higher internet  romantic  famrel  freetime goout Dalc Walc health  \
0   ...       yes       no        no       4         3     4    1    1      3   
1   ...       yes      yes        no       5         3     3    1    1      3   
2   ...       yes      yes        no       4         3     2    2    3      3   
3   ...       yes      yes       yes       3         2     2    1    1      5   
4   ...       yes       no        no       4         3     2    1    2      5   

  absences  
0        6  
1        4  
2       10  
3        2  
4        4  

[5 rows x 30 columns]

Preprocess Feature Columns

As you can see, there are several non-numeric columns that need to be converted! Many of them are simply yes/no, e.g. internet. These can be reasonably converted into 1/0 (binary) values.

Other columns, like Mjob and Fjob, have more than two values, and are known as categorical variables. The recommended way to handle such a column is to create as many columns as possible values (e.g. Fjob_teacher, Fjob_other, Fjob_services, etc.), and assign a 1 to one of them and 0 to all others.

These generated columns are sometimes called dummy variables, and we will use the pandas.get_dummies() function to perform this transformation. Run the code cell below to perform the preprocessing routine discussed in this section.


In [4]:
def preprocess_features(X):
    ''' Preprocesses the student data and converts non-numeric binary variables into
        binary (0/1) variables. Converts categorical variables into dummy variables. '''
    
    # Initialize new output DataFrame
    output = pd.DataFrame(index = X.index)

    # Investigate each feature column for the data
    for col, col_data in X.iteritems():
        
        # If data type is non-numeric, replace all yes/no values with 1/0
        if col_data.dtype == object:
            col_data = col_data.replace(['yes', 'no'], [1, 0])

        # If data type is categorical, convert to dummy variables
        if col_data.dtype == object:
            # Example: 'school' => 'school_GP' and 'school_MS'
            col_data = pd.get_dummies(col_data, prefix = col)  
        
        # Collect the revised columns
        output = output.join(col_data)
    
    return output

X_all = preprocess_features(X_all)
print "Processed feature columns ({} total features):\n{}".format(len(X_all.columns), list(X_all.columns))


Processed feature columns (48 total features):
['school_GP', 'school_MS', 'sex_F', 'sex_M', 'age', 'address_R', 'address_U', 'famsize_GT3', 'famsize_LE3', 'Pstatus_A', 'Pstatus_T', 'Medu', 'Fedu', 'Mjob_at_home', 'Mjob_health', 'Mjob_other', 'Mjob_services', 'Mjob_teacher', 'Fjob_at_home', 'Fjob_health', 'Fjob_other', 'Fjob_services', 'Fjob_teacher', 'reason_course', 'reason_home', 'reason_other', 'reason_reputation', 'guardian_father', 'guardian_mother', 'guardian_other', 'traveltime', 'studytime', 'failures', 'schoolsup', 'famsup', 'paid', 'activities', 'nursery', 'higher', 'internet', 'romantic', 'famrel', 'freetime', 'goout', 'Dalc', 'Walc', 'health', 'absences']

In [5]:
X_all.head()


Out[5]:
school_GP school_MS sex_F sex_M age address_R address_U famsize_GT3 famsize_LE3 Pstatus_A ... higher internet romantic famrel freetime goout Dalc Walc health absences
0 1.0 0.0 1.0 0.0 18 0.0 1.0 1.0 0.0 1.0 ... 1 0 0 4 3 4 1 1 3 6
1 1.0 0.0 1.0 0.0 17 0.0 1.0 1.0 0.0 0.0 ... 1 1 0 5 3 3 1 1 3 4
2 1.0 0.0 1.0 0.0 15 0.0 1.0 0.0 1.0 0.0 ... 1 1 0 4 3 2 2 3 3 10
3 1.0 0.0 1.0 0.0 15 0.0 1.0 1.0 0.0 0.0 ... 1 1 1 3 2 2 1 1 5 2
4 1.0 0.0 1.0 0.0 16 0.0 1.0 1.0 0.0 0.0 ... 1 0 0 4 3 2 1 2 5 4

5 rows × 48 columns

Implementation: Training and Testing Data Split

So far, we have converted all categorical features into numeric values. For the next step, we split the data (both features and corresponding labels) into training and test sets. In the following code cell below, you will need to implement the following:

  • Randomly shuffle and split the data (X_all, y_all) into training and testing subsets.
    • Use 300 training points (approximately 75%) and 95 testing points (approximately 25%).
    • Set a random_state for the function(s) you use, if provided.
    • Store the results in X_train, X_test, y_train, and y_test.

In [6]:
# Load train_test_split library
from sklearn.cross_validation import train_test_split

# Set the number of training points
num_train = 300

# Set the number of testing points
num_test = X_all.shape[0] - num_train

# Shuffle and split the dataset into the number of training and testing points above
X_train, X_test, y_train, y_test = train_test_split(X_all, y_all, test_size=num_test, random_state=0)

# Show the results of the split
print "Training set has {} samples.".format(X_train.shape[0])
print "Testing set has {} samples.".format(X_test.shape[0])


Training set has 300 samples.
Testing set has 95 samples.

Training and Evaluating Models

In this section, you will choose 3 supervised learning models that are appropriate for this problem and available in scikit-learn. You will first discuss the reasoning behind choosing these three models by considering what you know about the data and each model's strengths and weaknesses. You will then fit the model to varying sizes of training data (100 data points, 200 data points, and 300 data points) and measure the F1 score. You will need to produce three tables (one for each model) that shows the training set size, training time, prediction time, F1 score on the training set, and F1 score on the testing set.

The following supervised learning models are currently available in scikit-learn that you may choose from:

  • Gaussian Naive Bayes (GaussianNB)
  • Decision Trees
  • Ensemble Methods (Bagging, AdaBoost, Random Forest, Gradient Boosting)
  • K-Nearest Neighbors (KNeighbors)
  • Stochastic Gradient Descent (SGDC)
  • Support Vector Machines (SVM)
  • Logistic Regression

Self guided research regarding each learning model (pulled from Scikit documentation)

  • Gaussian Naive Bayes (GaussianNB)
    • Description: Set of algorithms based on applying Bayes' theorem with the "naive" assumption of independence between every pair of features.
    • Pros: In spite of their apparently over-simplified assumptions, naive Bayes classifiers have worked quite well in many real-world situations, famously document classification and spam filtering. They require a small amount of training data to estimate the necessary parameters. Naive Bayes learners and classifiers can be extremely fast compared to more sophisticated methods. The decoupling of the class conditional feature distributions means that each distribution can be independently estimated as a one dimensional distribution. This in turn helps to alleviate problems stemming from the curse of dimensionality.
    • Cons: On the flip side, although naive Bayes is known as a decent classifier, it is known to be a bad estimator.
    • Comment: Might work well, given the size of the small training dataset (300 observations).
  • Decision Trees
    • Description: Non-parametric supervised algorithm that predicts the value of a target variable by learning simple decision rules inferred from the data features.
    • Pros: Simple to understand and to interpret; trees can be visualised. Requires little data preparation. The cost of using the tree (i.e., predicting data) is logarithmic in the number of data points used to train the tree. Able to handle both numerical and categorical data. Able to handle multi-output problems. Performs well even if its assumptions are somewhat violated by the true model from which the data were generated.
    • Cons: Can create over-complex trees that overfit data (pruning is necessary). Small variations might cause large changes to the tree (mitigated through ensembling). Some relationships are hard to learn because trees cannot easily express them (XOR). Can be biased if some classes dominate.
    • Comment: Might not be ideal without an ensemble, and could also have difficultly in generalizing with small data set.
  • Ensemble Methods (Bagging, AdaBoost, Random Forest, Gradient Boosting)
    • Description: Encapsulates a large number of methods, but all built on the principle that by averaging many hypotheses regarding the dataset, the eventual algorithm produced will generalize better to future prediction problems.
    • Pros/Cons: Somewhat method specific (see description), but ensemble methods are incredibly powerful and have various uses in real-world applications.
    • Comment: Would be great to try one out.
  • K-Nearest Neighbors (KNeighbors)
    • Description: Algorithm works by finding a predefined number of training samples closest in distance to the new point, and predict the label from these ("k" nearest neighbors).
    • Pros: Very simple implementation and highly interpretable results (observations with similar attributes should behave similarly). Since it's a non-parametric method, works well for problems with irregular/non-linear decision boundaries.
    • Cons: For problems of high dimensionality, method is less effective due to "curse of dimensionality".
    • Comment: Likely not ideal given the high feature dimensionality, and relatively small training set.
  • Stochastic Gradient Descent (SGDC)
    • Description: a simple yet very efficient approach to discriminative learning of linear classifiers under convex loss functions such as (linear) Support Vector Machines and Logistic Regression.
    • Pros: Efficiency. Ease of implementation (lots of opportunities for code tuning).
    • Cons: SGD requires a number of hyperparameters such as the regularization parameter and the number of iterations. SGD is sensitive to feature scaling. Most successful for large-scale and spare machine learning problems (text classification and natural langauge processing); can easily scale beyond 10^5 observations and 10^5 features.
    • Comment: Likely not ideal for current problem, with significantly fewer observations and features.
  • Support Vector Machines (SVM)
    • Description: a discriminative classifier formally defined by a separating hyperplane. In other words, given labeled training data (supervised learning), the algorithm outputs an optimal hyperplane which categorizes new examples.
    • Pros: Effective in high dimensional spaces, even where feature space dimension > observation count. Uses a subset of training points in the decision function (called support vectors), so it is also memory efficient. Versatile: different Kernel functions can be specified for the decision function. Common kernels are provided, but it is also possible to specify custom kernels.
    • Cons: If the number of features is much greater than the number of samples, the method is likely to give poor performances. SVMs do not directly provide probability estimates.
    • Comment: Might be ideal given it's ability to handle problems where feature dim and observation count are comparable.
  • Logistic Regression
    • Description: a linear model for classification, it is also known as logit regression, maximum-entropy classification (MaxEnt) or the log-linear classifier. In this model, the probabilities describing the possible outcomes of a single trial are modeled using a logistic function.
    • Pros: Probability (likelihood of it meeting a conditional label) are returned, rather than discrete output, so thresholds can be tuned and set accordingly. It is also efficient in terms of time and memory requirement. It also handles small noise well.
    • Cons: Doesn’t perform well when feature space is too large. Doesn’t handle large number of categorical features/variables well. Relies on transformations for non-linear features.
    • Comment: Given its moderate ability to handle problems with larger feature dimensional spaces, it makes it less desireable to implement in the current context.

Question 2 - Model Application

List three supervised learning models that are appropriate for this problem. For each model chosen

  • Describe one real-world application in industry where the model can be applied. (You may need to do a small bit of research for this — give references!)
  • What are the strengths of the model; when does it perform well?
  • What are the weaknesses of the model; when does it perform poorly?
  • What makes this model a good candidate for the problem, given what you know about the data?

Answer:

Given some the cursory research performed previously, we've selected support vector machines, ensemble methods (of some type), and Gaussian naive Bayes classifiers. Answers to the above question are below.

  • Support Vector Machines
    • Real world application? Handwriting recognition; huge variability in observations makes it rather challenging, but researches have shown benefits of applying SVMs for these problems. See this article for an example.
    • Strengths? Effective in high dimensional spaces, even where feature space dimension > observation count. Uses a subset of training points in the decision function (called support vectors), so it is also memory efficient. Versatile: different Kernel functions can be specified for the decision function. Common kernels are provided, but it is also possible to specify custom kernels.
    • Weaknesses? If the number of features is much greater than the number of samples, the method is likely to give poor performances. SVMs do not directly provide probability estimates.
    • Rationale for selection? It seems to be quite good at problems where feature space and number of observations (training set) are comparable. Given the optionaly of various Kernel functions, it seems relatively flexible as well. Although its output is less than interpretable (as to understanding the reasons why future predictions occur as they do), it is not a requirement for our current purpose.
  • Ensemble Method (of some type)
    • Preliminaries Ensemble methods generally are classified as (1) averaging methods, e.g., bagging, random forest, where many estimators are independently built and averaged together to create a combined estimator or (2) boosting methods, e.g., AdaBoost, gradient boosting, where several "base" estimators are built and combined together to create the combined estimator. Although they appear similar, bagging methods attempt to reduce variance by averaging, while boosting methods reduce bias by creating complex estimators from much simplier ones. It's been noted that boosting methods generally have more hyper-parameters to tune and also more prone to overfitting. Due to the small number of observations, overfitting may be an issue, so bagging will be selected, Random Forest to be exact.
    • Real world application? Breast cancer detection, See this page for more information.
    • Strengths? For many datasets, it generalizes quite well. It runs efficiently with large datasets. It produces variable importance of the classifier. It handles missing data well, and can be tuned to handle unbalanced populations.
    • Weaknesses? Random forests have been observed to overfit for some datasets with noisy classification/regression tasks. Generally RFs are hard to interpret. Categorical data with several levels can provide difficultly for RFs. If the data contains groups of correlated features, then smaller groups are favored over larger groups.
    • Rationale for selection? While it isn't known how "noisy" this classification task is, it seems to handle a variety of problems quite well. Other downsides including model interpretability are not necessary for this classification exercise, although it is able to provide an estimate of "variable importance", which may be helpful for school administrators when developing proactive support (rather than reactive, as in this case).
  • Gaussian Naive Bayes
    • Real world application? Commonly used for spam email detection.
    • Strengths? In spite of their apparently over-simplified assumptions, naive Bayes classifiers have worked quite well in many real-world situations, famously document classification and spam filtering. They require a small amount of training data to estimate the necessary parameters. Naive Bayes learners and classifiers can be extremely fast compared to more sophisticated methods. The decoupling of the class conditional feature distributions means that each distribution can be independently estimated as a one dimensional distribution. This in turn helps to alleviate problems stemming from the curse of dimensionality.
    • Weaknesses? Although naive Bayes is known as a decent classifier, it is known to be a bad estimator. It also has strong feature independence assumptions (but doesn't cause too much trouble in typical applications).
    • Rationale for selection? The strengths seems to outweigh its weaknesses for this given problem. For example, avoiding the curse of dimensionality is a huge current benefit.

Setup

Run the code cell below to initialize three helper functions which you can use for training and testing the three supervised learning models you've chosen above. The functions are as follows:

  • train_classifier - takes as input a classifier and training data and fits the classifier to the data.
  • predict_labels - takes as input a fit classifier, features, and a target labeling and makes predictions using the F1 score.
  • train_predict - takes as input a classifier, and the training and testing data, and performs train_clasifier and predict_labels.
    • This function will report the F1 score for both the training and testing data separately.

In [7]:
def train_classifier(clf, X_train, y_train):
    ''' Fits a classifier to the training data. '''
    
    # Start the clock, train the classifier, then stop the clock
    start = time()
    clf.fit(X_train, y_train)
    end = time()
    
    # Print the results
    print "Trained model in {:.4f} seconds".format(end - start)

    
def predict_labels(clf, features, target):
    ''' Makes predictions using a fit classifier based on F1 score. '''
    
    # Start the clock, make predictions, then stop the clock
    start = time()
    y_pred = clf.predict(features)
    end = time()
    
    # Print and return results
    print "Made predictions in {:.4f} seconds.".format(end - start)
    return f1_score(target.values, y_pred, pos_label='yes')


def train_predict(clf, X_train, y_train, X_test, y_test):
    ''' Train and predict using a classifer based on F1 score. '''
    
    # Indicate the classifier and the training set size
    print "Training a {} using a training set size of {}. . .".format(clf.__class__.__name__, len(X_train))
    
    # Train the classifier
    train_classifier(clf, X_train, y_train)
    
    # Print the results of prediction for both training and testing
    print "F1 score for training set: {:.4f}.".format(predict_labels(clf, X_train, y_train))
    print "F1 score for test set: {:.4f}.".format(predict_labels(clf, X_test, y_test))

Implementation: Model Performance Metrics

With the predefined functions above, you will now import the three supervised learning models of your choice and run the train_predict function for each one. Remember that you will need to train and predict on each classifier for three different training set sizes: 100, 200, and 300. Hence, you should expect to have 9 different outputs below — 3 for each model using the varying training set sizes. In the following code cell, you will need to implement the following:

  • Import the three supervised learning models you've discussed in the previous section.
  • Initialize the three models and store them in clf_A, clf_B, and clf_C.
    • Use a random_state for each model you use, if provided.
    • Note: Use the default settings for each model — you will tune one specific model in a later section.
  • Create the different training set sizes to be used to train each model.
    • Do not reshuffle and resplit the data! The new training points should be drawn from X_train and y_train.
  • Fit each model with each training set size and make predictions on the test set (9 in total).
    Note: Three tables are provided after the following code cell which can be used to store your results.

In [8]:
# Import the three supervised learning models from sklearn
from sklearn import svm
from sklearn.ensemble import RandomForestClassifier
from sklearn.naive_bayes import GaussianNB

# Initialize the three models
seed = 1
clf_A = svm.SVC(random_state=seed)
clf_B = RandomForestClassifier(random_state=seed)
clf_C = GaussianNB()

# Train and evaluate classifiers
for size in [100, 200, 300]:
    print("\n")
    train_predict(clf_A, X_train.iloc[0:size], y_train.iloc[0:size], X_test, y_test)
    train_predict(clf_B, X_train.iloc[0:size], y_train.iloc[0:size], X_test, y_test)
    train_predict(clf_C, X_train.iloc[0:size], y_train.iloc[0:size], X_test, y_test)



Training a SVC using a training set size of 100. . .
Trained model in 0.0018 seconds
Made predictions in 0.0007 seconds.
F1 score for training set: 0.8591.
Made predictions in 0.0006 seconds.
F1 score for test set: 0.7838.
Training a RandomForestClassifier using a training set size of 100. . .
Trained model in 0.0229 seconds
Made predictions in 0.0010 seconds.
F1 score for training set: 0.9844.
Made predictions in 0.0007 seconds.
F1 score for test set: 0.7669.
Training a GaussianNB using a training set size of 100. . .
Trained model in 0.0006 seconds
Made predictions in 0.0003 seconds.
F1 score for training set: 0.8550.
Made predictions in 0.0002 seconds.
F1 score for test set: 0.7481.


Training a SVC using a training set size of 200. . .
Trained model in 0.0027 seconds
Made predictions in 0.0020 seconds.
F1 score for training set: 0.8693.
Made predictions in 0.0010 seconds.
F1 score for test set: 0.7755.
Training a RandomForestClassifier using a training set size of 200. . .
Trained model in 0.0219 seconds
Made predictions in 0.0011 seconds.
F1 score for training set: 0.9924.
Made predictions in 0.0007 seconds.
F1 score for test set: 0.7153.
Training a GaussianNB using a training set size of 200. . .
Trained model in 0.0007 seconds
Made predictions in 0.0003 seconds.
F1 score for training set: 0.8321.
Made predictions in 0.0005 seconds.
F1 score for test set: 0.7132.


Training a SVC using a training set size of 300. . .
Trained model in 0.0055 seconds
Made predictions in 0.0037 seconds.
F1 score for training set: 0.8692.
Made predictions in 0.0013 seconds.
F1 score for test set: 0.7586.
Training a RandomForestClassifier using a training set size of 300. . .
Trained model in 0.0244 seconds
Made predictions in 0.0013 seconds.
F1 score for training set: 0.9976.
Made predictions in 0.0008 seconds.
F1 score for test set: 0.7727.
Training a GaussianNB using a training set size of 300. . .
Trained model in 0.0008 seconds
Made predictions in 0.0003 seconds.
F1 score for training set: 0.8088.
Made predictions in 0.0003 seconds.
F1 score for test set: 0.7500.

Tabular Results

Edit the cell below to see how a table can be designed in Markdown. You can record your results from above in the tables provided.

Classifer 1 - Support Vector Machine

Training Set Size Training Time Prediction Time (test) F1 Score (train) F1 Score (test)
100 0.0011 0.0008 0.8591 0.7838
200 0.0028 0.0011 0.8693 0.7755
300 0.0052 0.0024 0.8692 0.7586

Classifer 2 - Random Forest

Training Set Size Training Time Prediction Time (test) F1 Score (train) F1 Score (test)
100 0.0219 0.0007 0.9844 0.7669
200 0.0231 0.0007 0.9924 0.7153
300 0.0201 0.0007 0.9976 0.7727

Classifer 3 - Gaussian Naive Bayes

Training Set Size Training Time Prediction Time (test) F1 Score (train) F1 Score (test)
100 0.0005 0.0002 0.8550 0.7481
200 0.0006 0.0002 0.8321 0.7132
300 0.0007 0.0002 0.8088 0.7500

Choosing the Best Model

In this final section, you will choose from the three supervised learning models the best model to use on the student data. You will then perform a grid search optimization for the model over the entire training set (X_train and y_train) by tuning at least one parameter to improve upon the untuned model's F1 score.

Question 3 - Choosing the Best Model

Based on the experiments you performed earlier, in one to two paragraphs, explain to the board of supervisors what single model you chose as the best model. Which model is generally the most appropriate based on the available data, limited resources, cost, and performance?

Answer: Based on available data, limited resources, cost, and performance, the Random Forest (RF) method appears to be the most appropriate. While it relatively more costly to train, by approximately 10x ~ 100x over the other two methods, at ~0.02s, for any training set size, this disadvantage is of little concern (especially considering that it remains roughly constant, regardless of training dataset size). Also, prediction time was blazing fast at ~0.0007s. Although all three classifiers demonstrated F1 accuracy near 75%, RF had F1 train scores of nearly unity. The larger disparity between F1 train and test scores seen for the RF method possibly suggests that it was overfitting the data and can be tuned to decrease such high variance.

Although beyond the scope of this project, the RF method can also identify the features most strongly associated with students' failure, which may allow school administrators to focus on proactive measures as well. Moreover, RFs are easier to explain to the general public, as decision trees, which are central to RFs, are quite intuitive.

Question 4 - Model in Layman's Terms

In one to two paragraphs, explain to the board of directors in layman's terms how the final model chosen is supposed to work. Be sure that you are describing the major qualities of the model, such as how the model is trained and how the model makes a prediction. Avoid using advanced mathematical or technical jargon, such as describing equations or discussing the algorithm implementation.

Answer: We have been able to leverage past student data, including attributes (e.g., gender, age, familial size, alcohol consumption) and performance (i.e., pass/fail final exam), to develop an algorithm that is able to classify future student failure/pass behavior for the high school final exam.

Using 300 previous observations, i.e., data from 300 students, including both attributes and performance, we were able to create a random forest model -- which sounds complicated, but isn't! First, it is helpful to understand that a random forest model is composed of decision trees -- the same kind of decision trees that your brain is creating and invoking every moment of each day. For example, when you start your day and decide if you should grab a coat, your mind may first ask: will the temperature remain above 72F? If yes, then secondly it may ask, is rain in the forecast? Maybe, if it's above 72F and sunny, then your brain will forego the coat. However, had it been less than 72F or rainy, then a coat would have accompanied your day. This process of leveraging two features, e.g., temperature, running a logical evaluation process, e.g., is it greater than 72F, and returning a result, e.g., grab a coat, is just one example of how our brains use decision trees daily.

A random forest model employs a similar approach, first creating a tree-at-a-time. In general, to create a tree from these variables, it attempts to find features (sunny/rainy, temperature) that divides the population into roughly equivalent samples. Essentially, it uses a divide and conquer approach, so if previous observations were all sunny, but had a good mix of temperature ranges, it would first use temperature to subdivide the population into smaller groups. This process of tree building is generally continued until the the tree's leafs reach desireablly small groups, or until all features are used. When the tree is complete, the leafs that contain observations are labelled according to the majority label (pass or fail). For example, if a leaf at the end of a tree contains 10 students that have passed, but 3 that have failed, it will label this "bucket" as "pass", such that when a future prediction occurs for another student that leads them into the same bucket, the model will return predict a "pass" as well.

However, the problem creating a single tree is that it could have built a model that was too specific to the training observations (we call this overfitting). To help the model generalize such that it'll perform well on future unseen data, the random forest method builds many trees and "averages" the results to better generalize to future data.

Implementation: Model Tuning

Fine tune the chosen model. Use grid search (GridSearchCV) with at least one important parameter tuned with at least 3 different values. You will need to use the entire training set for this. In the code cell below, you will need to implement the following:

  • Import sklearn.grid_search.GridSearchCV and sklearn.metrics.make_scorer.
  • Create a dictionary of parameters you wish to tune for the chosen model.
    • Example: parameters = {'parameter' : [list of values]}.
  • Initialize the classifier you've chosen and store it in clf.
  • Create the F1 scoring function using make_scorer and store it in f1_scorer.
    • Set the pos_label parameter to the correct value!
  • Perform grid search on the classifier clf using f1_scorer as the scoring method, and store it in grid_obj.
  • Fit the grid search object to the training data (X_train, y_train), and store it in grid_obj.

In [9]:
# Import 'GridSearchCV' and 'make_scorer'
from sklearn.grid_search import GridSearchCV
from sklearn.metrics import make_scorer

# Create the parameters list you wish to tune
parameters = [{
    "criterion": ["gini","entropy"],
    "max_depth": np.arange(4,8,2),
    "max_features": np.arange(5,9,2),
    "n_estimators": np.arange(11,15,2)
}]

# Initialize the classifier
seed = 1
clf = RandomForestClassifier(random_state=seed, n_jobs=1)

# Make an f1 scoring function using 'make_scorer' 
f1_scorer = make_scorer(f1_score, pos_label="yes")

# Perform grid search on the classifier using the f1_scorer as the scoring method
grid_obj = GridSearchCV(clf, parameters, scoring=f1_scorer)

# TODO: Fit the grid search object to the training data and find the optimal parameters
grid_obj.fit(X_train, y_train)

# Get the estimator
clf = grid_obj.best_estimator_
print grid_obj.best_params_
print grid_obj.best_score_

# Report the final F1 score for training and testing after parameter tuning
print "Tuned model has a training F1 score of {:.4f}.".format(predict_labels(clf, X_train, y_train))
print "Tuned model has a testing F1 score of {:.4f}.".format(predict_labels(clf, X_test, y_test))


{'max_features': 7, 'n_estimators': 13, 'criterion': 'gini', 'max_depth': 6}
0.822971961345
Made predictions in 0.0014 seconds.
Tuned model has a training F1 score of 0.9031.
Made predictions in 0.0010 seconds.
Tuned model has a testing F1 score of 0.8058.

Question 5 - Final F1 Score

What is the final model's F1 score for training and testing? How does that score compare to the untuned model?

Answer: A handful of numerical experiments demonstrated some rather interesting results.

Four variables were considered: criterion, max_depth, max_features, and n_estimators. An iterative mutltigrid-like approach was employed, first beginning with large coarse grids, and slowly switching to smaller, finer grids, once appropriate ranges for each variable were empirically determined. This was done to lessen the overall compuational burden of the grid search method.

Using step sizes of 1, the following parametrized values were found to be optimal, according to the grid search: {'max_features': 6, 'n_estimators': 11, 'criterion': 'gini', 'max_depth': 6} This resulted in a grid search score of 0.83, but when these parameters were used to create the 'tuned' model, the F1 training and test scores were 0.9095 and 0.7639. Unexpectedly, the test score was slightly less than the 'untuned' model test score.

Inquisitive, upon reverting back to step sizes of 2, the following parametrized values were found to be optimal: {'max_features': 7, 'n_estimators': 13, 'criterion': 'gini', 'max_depth': 6} Here, the difference being that _maxfeatures was bumped from 6 to 7. As expected, reducing the resolution obtainable by the grid search, the resulting score also reduced -- slightly from 0.83 to 0.8229. However, by under optimizing the grid search result, these new parameters created 'tuned' model F1 train and test scores of 0.9031 and 0.8058.

Fascinating!

There are two things likely at play here. First, is that the grid search metric for optimal parameters and the test set metric are dissimilar. The grid search metric is using a k-fold cross validation style on training data, while the test metric is computing an F1 score for a test set. Since the populations and methodologies are different, we'd not to surprised by these dissimilar results. Additionally, the small sample size is likely also reducing the model's ability to generalize. Perhaps, had more observations been available, these differences would become less significant.

Note: Once you have completed all of the code implementations and successfully answered each question above, you may finalize your work by exporting the iPython Notebook as an HTML document. You can do this by using the menu above and navigating to
File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.