def buildAndTestModel():
houseTrain, inputCols, outputCol = reading.readData("data/train.csv")
inputCols.remove('Id')
houseTest = reading.readTestData("data/test.csv")
preprocessing.manageNAValues(houseTrain, inputCols)
preprocessing.preprocess(houseTest, houseTrain, inputCols)
preprocessing.preprocess(houseTrain, houseTrain, inputCols)
alg = GradientBoostingRegressor(random_state=1, n_estimators=800)
alg.fit(houseTrain.loc[:, inputCols], houseTrain.loc[:, outputCol])
predictions = alg.predict(houseTest.loc[:, inputCols])
submitDF = pd.DataFrame({
"Id": houseTest.loc[:, "Id"],
"SalePrice": predictions
})
submitDF.to_csv("data/submission.csv", index=False)
def buildAndTestModel():
houseTrain, inputCols, outputCol = reading.readData()
preprocessing.preprocess(houseTrain, houseTrain, inputCols)
alg = GradientBoostingRegressor(random_state=1, n_estimators=800)
cvScores = model_selection.cross_val_score(alg,
houseTrain.loc[:, inputCols],
houseTrain.loc[:, outputCol],
cv=10,
scoring='r2')
print(np.mean(cvScores))
houseTest = pd.read_csv("data/test.csv")
preprocessing.preprocess(houseTest, houseTrain, inputCols)
alg = GradientBoostingRegressor(random_state=1, n_estimators=800)
alg.fit(houseTrain.loc[:, inputCols], houseTrain.loc[:, outputCol])
predictions = alg.predict(houseTest.loc[:, inputCols])
submitDF = pd.DataFrame({
"Row": houseTest.loc[:, "SalePrice"],
"Prediction": predictions
})
submitDF.to_csv("data/submission.csv", index=False)
def main():
trainDF, inputsCol, outputCol = reading.readData()
preprocessing.preprocess(trainDF, trainDF, inputsCol)
alg = GradientBoostingRegressor(
random_state=1, n_estimators=800
) # accuracy does not change everytime it is run with set random_state
# additionalCols = ['Attic', 'Finished', 'Split', 'Foyer', 'Duplex', 'Pud', 'Conversion', 'Story']
# inputsCol = inputsCol + additionalCols
# inputsCol = list(set(inputsCol) - set(['MSSubClass']))
cvScores = model_selection.cross_val_score(alg,
trainDF.loc[:, inputsCol],
trainDF.loc[:, outputCol],
cv=10,
scoring='r2')
print("Highest Accuracy with all features, default parameterizations =",
np.mean(cvScores))
# inputsCol = additionalCols + [outputCol]
# visualization.visualize(trainDF, additionalCols, outputCol)
#visualization.visualize(trainDF,inputsCol,outputCol)
#Already done, testing what happens to accuracy removing one feature at a time
"""
inputsColTemp = copy.deepcopy(inputsCol)
temp = {}
temp["Nothing removed"] = np.mean(cvScores)
while len(inputsColTemp) != 0:
featureRemoved = inputsColTemp.pop()
inputsCol.remove(featureRemoved)
alg = GradientBoostingRegressor(random_state = 1, n_estimators = 800)
cvScores = model_selection.cross_val_score(alg, trainDF.loc[:, inputsCol], trainDF.loc[:, outputCol], cv=10, scoring='r2')
temp[featureRemoved] = np.mean(cvScores)
print("Accuracy when removing " + featureRemoved + " =", np.mean(cvScores))
inputsCol.append(featureRemoved)
export = pd.Series(temp)
export.to_csv(os.getcwd() + '/removeOneFeature_postParameterization.csv')
"""
#Testing various parameterizations
'''
print("Changing n:")
i = 100
while i<=1000:
alg = GradientBoostingRegressor(random_state = 1, n_estimators = i)
cvScores = model_selection.cross_val_score(alg, trainDF.loc[:, inputsCol], trainDF.loc[:, outputCol], cv=10, scoring='r2')
print("For score of i = " + str(i) + ":", np.mean(cvScores))
i += 100
'''
'''
print("Changing learning rate:")
j = 0.01
while j<=1:
alg = GradientBoostingRegressor(random_state = 1, learning_rate = j)
cvScores = model_selection.cross_val_score(alg, trainDF.loc[:, inputsCol], trainDF.loc[:, outputCol], cv=10, scoring='r2')
print("For rate of j = " + str(j) + ":", np.mean(cvScores))
j += 0.05
'''
"""
print("Testing proposed optimum settings (approximate):")
alg = GradientBoostingRegressor(random_state = 1, n_estimators = 900, learning_rate = 0.16)
cvScores = model_selection.cross_val_score(alg, trainDF.loc[:, inputsCol], trainDF.loc[:, outputCol], cv=10, scoring='r2')
print("Score:", np.mean(cvScores))
"""
"""
For score of i = 100: 0.9695838362322334
For score of i = 200: 0.9713926415403247
For score of i = 300: 0.97218152251929
For score of i = 400: 0.9725433964858627
For score of i = 500: 0.9727129864730497
For score of i = 600: 0.9727150286009545
For score of i = 700: 0.9727550478839996
For score of i = 800: 0.9728148128730638
For score of i = 900: 0.9727879075157484
For score of i = 1000: 0.9727582470818399
"""
"""
Changing learning rate:
For rate of j = 0.01: 0.7866244576707975
For rate of j = 0.060000000000000005: 0.9673647242443091
For rate of j = 0.11000000000000001: 0.9685577086455324
For rate of j = 0.16000000000000003: 0.9706671931416834
For rate of j = 0.21000000000000002: 0.9704141690215575
For rate of j = 0.26: 0.9693811639251051
For rate of j = 0.31: 0.9640193537044937
For rate of j = 0.36: 0.9662721202679398
For rate of j = 0.41: 0.9650110713531361
For rate of j = 0.45999999999999996: 0.9598916595713343
For rate of j = 0.51: 0.9612472671240511
For rate of j = 0.56: 0.9582289549106051
For rate of j = 0.6100000000000001: 0.9532347135301027
For rate of j = 0.6600000000000001: 0.9568633720893616
For rate of j = 0.7100000000000002: 0.951672708291688
For rate of j = 0.7600000000000002: 0.9524549341492469
For rate of j = 0.8100000000000003: 0.9461558271946371
For rate of j = 0.8600000000000003: 0.9354096980604721
For rate of j = 0.9100000000000004: 0.9328281792809602
For rate of j = 0.9600000000000004: 0.9325540177700384
"""
#Optimum for both at the same time = 0.9708920005637681... interesting... so looks like better to focus on n_estimators
#Testing combinations of removing items
# without ["HalfBath", "LandSlope", "BldgType", "YearBuilt", "LowQualFinSF", "Utilities"] 0.9697767256899044
# without ["HalfBath", "LandSlope", "BldgType", "YearBuilt", "LowQualFinSF"] 0.9689720298816091
# without ["HalfBath", "LandSlope", "BldgType", "YearBuilt"] 0.9689965270724571
# without ["HalfBath", "LandSlope", "BldgType"] 0.9701461139168301
# without ["HalfBath", "LandSlope"] 0.968366504937428
# without ["HalfBath"] 0.968366504937428
# TODO It seems even when deleting one column improves the result, removing the multiple columns worsen the accuracy
#Test function for above accuracies
#itemsToRemove = set(["HalfBath"]) # changed to ScreenPorch from ScreenProch
# TODO Why does the result different from the result in the while loop above? This should match up with the accuracy 0.97021332160605. Does the order of list matter??
#post_featureRemoval = filter(lambda x: x not in itemsToRemove, inputsCol)
# post_featureRemoval = list(set(inputsCol) - itemsToRemove)
# inputsCol.remove('HalfBath')
#alg = GradientBoostingRegressor(random_state = 1)
#cvScores = model_selection.cross_val_score(alg, trainDF.loc[:, post_featureRemoval], trainDF.loc[:, outputCol], cv=10, scoring='r2')
#print("After removing all detrimental features =", np.mean(cvScores))
# This results in a lower value... does this mean some of these are related, or simply need preprocessing
# (e.g., year built should probably become age)?
#visualization.visualize(trainDF,inputsCol,outputCol)
#Testing a different model (linear regression)
"""
def visualizationTest():
trainDF, inputsCol, outputCol = reading.readData()
inputsCol = ['MSSubClass']
visualization.visualizeScatterplot(trainDF, inputsCol, outputCol)