Education Data Mining Testing System

We need an in-house testing system to validate our machine learning algorithm. We need this in order to iterate towards better solutions. I am basing this in-house testing system on the Yu et al. JMLR Workshop and Conference Proceedings paper that the winning team submitted. The leaderboard contains the full list of submissions and links to papers.

In the Yu et al. paper, the main reason why they built their own testing system instead of just submitting to their answers and having the KDD Cup server score it was to avoid overfitting the solution.

In [1]:
%matplotlib inline
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import sklearn

Feature engineering

In [73]:
# Get the data: Algebra 2005-2006 (A56) and/or Algebra 2008-2009 (A89)
a56_train_filepath = 'data/algebra0506/algebra_2005_2006_train.txt'
#a89_train_filepath = 'data/algebra0809/algebra_2008_2009_train.txt'

a56data = pd.read_table(a56_train_filepath)
#a89data = pd.read_table(a89_train_filepath)
In [74]:
hierarchy = a56data['Problem Hierarchy']

Split the problem hierarchy into 'Units' and 'Sections'

In [75]:
units, sections = [], []
for i in range(len(hierarchy)):
    units.append(hierarchy[i].split(',')[0].strip())
    sections.append(hierarchy[i].split(',')[1].strip())
In [76]:
# Now add 'Units' and 'Sections' as columns within the dataframe
a56data['Problem Unit'] = pd.Series(units, index=a56data.index)
a56data['Problem Section'] = pd.Series(sections, index=a56data.index)
In [77]:
# Rearrange order of columns
cols = a56data.columns.tolist()
cols = cols[0:3]+cols[-2::]+cols[3:-2]
a56data = a56data[cols]

Create a temporary dataframe for the addition of new binary features

In [79]:
df = a56data

Split students into array of binary features

In [80]:
students = set(a56data['Anon Student Id'])
print 'There are {0} students, so we will be adding as many columns to this dataframe.'.format(len(students))
There are 574 students, so we will be adding as many columns to this dataframe.
In [81]:
numrows = len(df)

# Add a column for every student and fill them with zeros
for stud in students:
    df[stud] = pd.Series(np.zeros(numrows), index=df.index)
    
# For each student's problem entries in the dataframe, mark their column as 1
for stud in students:
    df.loc[df['Anon Student Id'] == stud,stud] = 1
In [82]:
np.shape(df)
Out[82]:
(809694, 595)

Split problem units into array of binary features

In [83]:
units = set(a56data['Problem Unit'])
print 'There are {0} unique problem units, so we will be adding as many columns to this dataframe.'.format(len(units))
There are 32 unique problem units, so we will be adding as many columns to this dataframe.
In [84]:
numrows = len(df)

# Add a column for every problem unit and fill them with zeros
for u in units:
    df[u] = pd.Series(np.zeros(numrows), index=df.index)
    
# For each student's attempt at a problem, mark the appropriate problem unit as 1
for u in units:
    df.loc[df['Problem Unit'] == u,u] = 1
In [85]:
np.shape(df)
Out[85]:
(809694, 627)

Split problem section into array of binary features

In [86]:
sections = set(a56data['Problem Section'])
print 'There are {0} unique problem sections, so we will be adding as many columns to this dataframe.'.format(len(sections))
There are 138 unique problem sections, so we will be adding as many columns to this dataframe.
In [87]:
numrows = len(df)

# Add a column for every problem unit and fill them with zeros
for s in sections:
    df[s] = pd.Series(np.zeros(numrows), index=df.index)
    
# For each student's attempt at a problem, mark the appropriate problem unit as 1
for s in sections:
    df.loc[df['Problem Section'] == s,s] = 1
In [88]:
np.shape(df)
Out[88]:
(809694, 765)

Split problem name into array of binary features

In [105]:
pnames = set(a56data['Problem Name'])
print 'There are {0} unique problem names, so we will be adding as many columns to this dataframe.'.format(len(pnames))
There are 1084 unique problem names, so we will be adding as many columns to this dataframe.
In [90]:
numrows = len(df)

# Add a column for every problem unit and fill them with zeros
for n in pnames:
    df[n] = pd.Series(np.zeros(numrows), index=df.index)
    
# For each student's attempt at a problem, mark the appropriate problem unit as 1
for n in pnames:
    df.loc[df['Problem Name'] == n,n] = 1
In [91]:
np.shape(df)
Out[91]:
(809694, 1849)

Split step name into array of binary features

In [106]:
snames = set(a56data['Step Name'])
print 'There are {0} unique step names, so we will be adding as many columns to this dataframe.'.format(len(snames))
There are 187539 unique step names, so we will be adding as many columns to this dataframe.
In [54]:
numrows = len(df)

# Add a column for every problem unit and fill them with zeros
for n in snames:
    df[n] = pd.Series(np.zeros(numrows), index=df.index)
    
# For each student's attempt at a problem, mark the appropriate problem unit as 1
for n in snames:
    df.loc[df['Step Name'] == n,n] = 1

Create the testing dataframe

In [92]:
# Create an empty testing dataframe
testdf = pd.DataFrame(columns=df.columns)

# Create the testing set
for i in range(len(unique_units)):
    # Get the last problem of the current problem unit
    lastProb = list(df[df['Problem Unit'] == unique_units[i]]['Problem Name'])[-1]
    
    # Get all the rows corresponding to the last problem for the given problem unit
    lastProbRows = a56data[(df['Problem Unit'] == unique_units[i]) & (df['Problem Name']==lastProb)]
    
    # Concatenate test dataframe with the rows just found
    testdf = pd.concat([testdf,lastProbRows])
In [93]:
# Create a training dataframe that is equal to original dataframe with all the test cases removed
trainIndex = df.index - testdf.index
traindf = df.loc[trainIndex]
In [94]:
# Get the target feature within the test set: the Correct First Attmpt
CFAs = np.array(testdf['Correct First Attempt'])

Test functions

In [95]:
# Define a helper function for calculating the root-mean-square error
def RMSE(p,y):
    ''' The Root-Mean-Square Error takes the predicted values p for the target
    variable y and takes the square root of the mean of the square of their
    differences. '''
    return np.sqrt(np.sum(np.square(p-y))/len(y))
In [96]:
# Test the RMSE for an array of all zeros
p = np.zeros(len(CFAs))
print 'An array of all zeros gives an RMSE of:',RMSE(p,CFAs)

# Test the RMSE for an array of all ones
p = np.ones(len(CFAs))
print 'An array of all ones gives an RMSE of:',RMSE(p,CFAs)

# Test the RMSE for an array of random 0s and 1s
p = np.random.randint(0,2,len(CFAs)).astype(float)
print 'An array of random ones and zeros gives an RMSE of:',RMSE(p,CFAs)
An array of all zeros gives an RMSE of: 0.863841709437
An array of all ones gives an RMSE of: 0.503763338322
An array of random ones and zeros gives an RMSE of: 0.710079831105
In [97]:
# Define the logistic function
def logisfunc(x):
    return 1.0 / (1.0 + np.exp(-x))
In [98]:
def logit_plots(X,y,model,feat_names):
    num_samples  = np.shape(X)[0]
    num_features = np.shape(X)[1]
    print num_samples, 'number of samples'
    print num_features, 'number of features'
    
    # The coefficients and bias of the decision function
    coefs = model.coef_.ravel()
    bias  = model.intercept_
    
    # Plot the decision function separately for each feature
    x_plt = np.linspace(-1.0,1.0,300)
    
    fig = plt.figure(figsize=(3*num_features,6))
    
    for feat in range(num_features):
        x     = x_plt * coefs[feat] + bias
        decs  = logisfunc(x)
        
        fig.add_subplot(num_features,1,feat+1)
        plt.plot(X[:,feat],y,'x')
        plt.plot(x_plt,decs)
        plt.axis((0,1,0,1))
        plt.xlabel(feat_names[feat])
    plt.tight_layout()
In [99]:
def error_metrics(p,yy):
    '''Calculates the error metrics, i.e. the precision and recall.
    Precision = True positives / Predicted positives
    Recall    = True positives / Actual positives'''
    predicted_positives = len(p[p==1])
    actual_positives    = len(yy[yy==1])
    # The predicted values for when actual values are 1
    pp = p[yy==1]
    # True positives are when these predicted values are also 1
    true_positives      = len(pp[pp==1])
    false_positives     = len(yy) - true_positives
    
    precision = float(true_positives) / float(predicted_positives)
    recall    = float(true_positives) / float(actual_positives)
    
    F_1score  = 2.0 * precision * recall / (precision + recall)
    
    print 'Root-mean-square error: ', RMSE(p,yy)
    
    print '\nPrecision: Of all predicted CFAs, what fraction actually succeeded?'
    print precision
    
    print '\nRecall: Of all actual CFAs, what fraction did we predict correctly?'
    print recall
    
    print '\nF_1 Score: ', F_1score

Machine learning

In [100]:
traindf.columns
Out[100]:
Index([u'Row', u'Anon Student Id', u'Problem Hierarchy', u'Problem Unit', u'Problem Section', u'Problem Name', u'Problem View', u'Step Name', u'Step Start Time', u'First Transaction Time', u'Correct Transaction Time', u'Step End Time', u'Step Duration (sec)', u'Correct Step Duration (sec)', u'Error Step Duration (sec)', u'Correct First Attempt', u'Incorrects', u'Hints', u'Corrects', u'KC(Default)', u'Opportunity(Default)', u'OCmTg4ha6w', u'M7hLaVJGvX', u'6bJ4auIa8L', u'D08sSQFrq3', u'J76B3lyQG9', u'77y7iIIXuv', u'k37lfl6ENf', u'olyVui8lHe', u'verm1Wp12u', u'S7E3jvZbGG', u'cua2tdf6zb', u'3Whv9UbPsR', u'3EWyQLUo83', u'XqBaV46VbC', u'qN6WN7C097', u'm9a501e1MM', u'8FcKH1d6A2', u'8eqtD31y66', u'7y5NJZ7S6u', u'rA9d62o2bb', u'5gtqSDwEt1', u'6d2wZ1x370', u'oeTCjHG37z', u'HvghFVBS5v', u'1to8tgu4IT', u'UO70mRwd30', u'XrH5AAPdtj', u'TAw598sUia', u'pI9Tpj13ty', u'v52MhQAVT9', u'nybp7zY98z', u'yuB6nxh3FX', u'bt0ZuHaCGY', u'dD5k5322J5', u'2AMmebFl86', u'45euP7C062', u'QS4cvQ8w0o', u'Dmq6441349', u'NX8N2fJ630', u'9hs21hUG5O', u'a47O44klh3', u'45TTYcotWk', u'U50h3ZFmGt', u'Vu26QoCms4', u'WmIXxzmmD3', u'6jCygPdswz', u'SH1cSNDIA8', u'eD0u9WOep7', u'X5eS5kvC8h', u'X223hIsDU4', u'Vc0J7iVot4', u'c8Gl35DPn4', u'EfN583275t', u'MPqJdqnPtc', u'hwF4tyWU50', u'0U9x5pNv8t', u'5x5fHvFFLv', u'RwM69ocq8e', u'DZXDgO0B3u', u'5ddRYL0LBA', u'4ajdr1a4Kc', u'st945Ucdi6', u'D431zXkvuC', u'ajPRq0Q08b', u'b2W4ZO3uLV', u'NZJd9Aq3De', u'3hSu07XfGd', u'80nlN05JQ6', u'z2zuhnARi6', u'w0FMzJORlK', u'ey9rvMnU57', u'XNEgfvXx50', u'On0ILT6rt5', u'XGFd357i6u', u'XHm4u4Y8OU', u'jx1ndu85oq', u'dc35t0IQHq', u'8r7mHkuKX9', u'c6LS9kDJe1', ...], dtype='object')
In [101]:
# Define a helper function to normalize the feature matrix X
import numba
def autonorm(X):
    ''' Calculates the mean and range of values of each column
    in the matrix (features) subtracts the mean from each value
    and divides by the range, thereby normalizing all values to
    fall between -1 and 1.'''
    x_means = np.mean(X,axis=0)
    x_means = np.ones(np.shape(X))*x_means
    x_maxs  = np.max(X,axis=0)
    x_mins  = np.min(X,axis=0)
    x_range = x_maxs - x_mins
    X_normd = (X - x_means) / x_range
    return X_normd

autonorm_jit = numba.jit(autonorm)
In [113]:
features_to_norm = ['Step Duration (sec)','Hints','Problem View']
binary_features = list(students)+list(units)+list(sections)+list(pnames)#+list(snames)
target_feature = ['Correct First Attempt']
all_features = features_to_norm + binary_features + target_feature
In [ ]:
X = traindf[all_features].dropna()
y = np.array(X[target_feature]).astype(int).ravel()
X_to_norm = np.array(X[features_to_norm])
X_nonnorm = np.array(X[binary_features])
X_to_norm = autonorm(X_to_norm)
X = np.concatenate((X_to_norm,X_nonnorm), axis=1)
In [104]:
XX = testdf[all_features].dropna()
yy = np.array(XX[target_feature]).astype(int).ravel()
XX_to_norm = np.array(XX[features_to_norm])
XX_nonnorm = np.array(XX[binary_features])
XX_to_norm = autonorm(XX_to_norm)
XX = np.concatenate((XX_to_norm,XX_nonnorm), axis=1)
In [25]:
from sklearn import linear_model
model = linear_model.LogisticRegression()
In [26]:
np.shape(X)
Out[26]:
(760650, 577)
In [27]:
model.fit(X,y)
Out[27]:
LogisticRegression(C=1.0, class_weight=None, dual=False, fit_intercept=True,
          intercept_scaling=1, penalty='l2', random_state=None, tol=0.0001)
In [28]:
p = model.predict(XX).astype(float)
In [30]:
error_metrics(p,yy)
Root-mean-square error:  0.401917482001

Precision: Of all predicted CFAs, what fraction actually succeeded?
0.831047542873

Recall: Of all actual CFAs, what fraction did we predict correctly?
0.983518472118

F_1 Score:  0.900877237721
In [32]:
np.shape(X)
Out[32]:
(760650, 577)
In [33]:
len(students)
Out[33]:
574
In [34]:
traindf.columns
Out[34]:
Index([u'Row', u'Anon Student Id', u'Problem Hierarchy', u'Problem Unit', u'Problem Section', u'Problem Name', u'Problem View', u'Step Name', u'Step Start Time', u'First Transaction Time', u'Correct Transaction Time', u'Step End Time', u'Step Duration (sec)', u'Correct Step Duration (sec)', u'Error Step Duration (sec)', u'Correct First Attempt', u'Incorrects', u'Hints', u'Corrects', u'KC(Default)', u'Opportunity(Default)', u'OCmTg4ha6w', u'M7hLaVJGvX', u'6bJ4auIa8L', u'D08sSQFrq3', u'J76B3lyQG9', u'77y7iIIXuv', u'k37lfl6ENf', u'olyVui8lHe', u'verm1Wp12u', u'S7E3jvZbGG', u'cua2tdf6zb', u'3Whv9UbPsR', u'3EWyQLUo83', u'XqBaV46VbC', u'qN6WN7C097', u'm9a501e1MM', u'8FcKH1d6A2', u'8eqtD31y66', u'7y5NJZ7S6u', u'rA9d62o2bb', u'5gtqSDwEt1', u'6d2wZ1x370', u'oeTCjHG37z', u'HvghFVBS5v', u'1to8tgu4IT', u'UO70mRwd30', u'XrH5AAPdtj', u'TAw598sUia', u'pI9Tpj13ty', u'v52MhQAVT9', u'nybp7zY98z', u'yuB6nxh3FX', u'bt0ZuHaCGY', u'dD5k5322J5', u'2AMmebFl86', u'45euP7C062', u'QS4cvQ8w0o', u'Dmq6441349', u'NX8N2fJ630', u'9hs21hUG5O', u'a47O44klh3', u'45TTYcotWk', u'U50h3ZFmGt', u'Vu26QoCms4', u'WmIXxzmmD3', u'6jCygPdswz', u'SH1cSNDIA8', u'eD0u9WOep7', u'X5eS5kvC8h', u'X223hIsDU4', u'Vc0J7iVot4', u'c8Gl35DPn4', u'EfN583275t', u'MPqJdqnPtc', u'hwF4tyWU50', u'0U9x5pNv8t', u'5x5fHvFFLv', u'RwM69ocq8e', u'DZXDgO0B3u', u'5ddRYL0LBA', u'4ajdr1a4Kc', u'st945Ucdi6', u'D431zXkvuC', u'ajPRq0Q08b', u'b2W4ZO3uLV', u'NZJd9Aq3De', u'3hSu07XfGd', u'80nlN05JQ6', u'z2zuhnARi6', u'w0FMzJORlK', u'ey9rvMnU57', u'XNEgfvXx50', u'On0ILT6rt5', u'XGFd357i6u', u'XHm4u4Y8OU', u'jx1ndu85oq', u'dc35t0IQHq', u'8r7mHkuKX9', u'c6LS9kDJe1', ...], dtype='object')
In [37]:
len(set(traindf['Step Name']))
Out[37]:
171886
In [38]:
len(traindf.columns)
Out[38]:
595
In [45]:
params = model.get_params(deep=True)
In [46]:
params
Out[46]:
{'C': 1.0,
 'class_weight': None,
 'dual': False,
 'fit_intercept': True,
 'intercept_scaling': 1,
 'penalty': 'l2',
 'random_state': None,
 'tol': 0.0001}
In [48]:
T = model.predict_proba(X)
In [51]:
T
Out[51]:
array([[ 0.70539268,  0.29460732],
       [ 0.18470565,  0.81529435],
       [ 0.45739163,  0.54260837],
       ..., 
       [ 0.17350595,  0.82649405],
       [ 0.70812947,  0.29187053],
       [ 0.14044815,  0.85955185]])
In [ ]: