Practical Predictive Modeling in Python

Practical Predictive
Modeling in Python
Robert Dempsey
robertwdempsey.com

Robert Dempsey
robertwdempsey
rdempsey
rdempsey
robertwdempsey.com

Doing All Things In SQL
Makes Panda sad and confused

Each New Thing You Learn
Leads to another new thing to learn, and another, and…

So Many Things
1. Which predictive modeling technique to use
2. How to get the data into a format for modeling
3. How to ensure the “right” data is being used
4. How to feed the data into the model
5. How to validate the model results
6. How to save the model to use in production
7. How to implement the model in production and apply it to new observations
8. How to save the new predictions
9. How to ensure, over time, that the model is correctly predicting outcomes
10.How to later update the model with new training data

Practical Predictive Modeling in Python

Model Selection
• How much data do you have?
• Are you predicting a category? A quantity?
• Do you have labeled data?
• Do you know the number of categories?
• How much data do you have?

Regression
• Used for estimating the relationships among
variables
• Use when:
• Predicting a quantity
• More than 50 samples

Classiﬁcation
• Used to answer “what is this object”
• Use when:
• Predicting a category
• Have labeled data

Clustering
• Used to group similar objects
• Use when:
• Predicting a category
• Don’t have labeled data
• Number of categories is known or unknown
• Have more than 50 samples

Dimensionality Reduction
• Process for reducing the number of random
variables under consideration (feature selection
and feature extraction)
• Use when:
• Not predicting a category or a quantity
• Just looking around

Model Selection
http://scikit-learn.org/stable/tutorial/machine_learning_map/

Format The Data
• Pandas FTW!
• Use the map() function to convert any text to a
number
• Fill in any missing values
• Split the data into features (the data) and targets
(the outcome to predict) using .values on the
DataFrame

map()
def update_failure_explanations(type):
if type == 'dob':
return 0
elif type == 'name':
return 1
elif type == 'ssn dob name':
return 2
elif type == 'ssn':
return 3
elif type == 'ssn name':
return 4
elif type == 'ssn dob':
return 5
elif type == 'dob name':
return 6

Fill In Missing Values
df.my_field.fillna(‘Missing', inplace=True)
df.fillna(0, inplace=True)

Split the Data
t_data = raw_data.iloc[:,0:22].values
1. Create a matrix of values
t_targets = raw_data['veriﬁed'].values
2. Create a matrix of targets

Get The Right Data
• This is called “Feature selection”
• Univariate feature selection
• SelectKBest removes all but the k highest scoring features
• SelectPercentile removes all but a user-speciﬁed highest scoring
percentage of features using common univariate statistical tests for
each feature: false positive rate
• SelectFpr, false discovery rate SelectFdr, or family wise error SelectFwe.
• GenericUnivariateSelect allows to perform univariate feature selection
with a conﬁgurable strategy.
http://scikit-learn.org/stable/modules/feature_selection.html

Data => Model
1. Build the model
http://scikit-learn.org/stable/modules/cross_validation.html
from sklearn import linear_model
logClassifier = linear_model.LogisticRegression(C=1,
random_state=111)
2. Train the model
from sklearn import cross_validation
X_train, X_test, y_train, y_test = cross_validation.train_test_split(the_data,
the_targets,
cv=12,
test_size=0.20,
random_state=111)
logClassifier.fit(X_train, y_train)

Validation
1. Accuracy Score
http://scikit-learn.org/stable/modules/cross_validation.html
from sklearn import metrics
metrics.accuracy_score(y_test, predicted)
2. Confusion Matrix
metrics.confusion_matrix(y_test, predicted)

Save the Model
Pickle it!
https://docs.python.org/3/library/pickle.html
import pickle
model_file = "/lr_classifier_09.29.15.dat"
pickle.dump(logClassifier, open(model_file, "wb"))
Did it work?
logClassifier2 = pickle.load(open(model, "rb"))
print(logClassifier2)

Implement in Production
• Clean the data the same way you did for the model
• Feature mappings
• Column re-ordering
• Create a function that returns the prediction
• Deserialize the model from the ﬁle you created
• Feed the model the data in the same order
• Call .predict() and get your answer

Example
def verify_record(record_scores):
# Reload the trained model
tif = "models/t_lr_classifier_07.28.15.dat"
log_classifier = pickle.load(open(tcf, "rb"))
# Return the prediction
return log_classifier.predict(record_scores)[0]

Save Your Predictions
As you would any other piece of data

Unleash the minion army!
… or get more creative

Be Smart
Train it again, but with validated predictions

Step Review
1. Select a predictive modeling technique to use
2. Get the data into a format for modeling
3. Ensure the “right” data is being used
4. Feed the data into the model
5. Validate the model results

Step Review
6. Save the model to use in production
7. Implement the model in production and apply it to
new observations
8. Save the new predictions
9. Ensure the model is correctly predicting outcomes
over time
10. Update the model with new training data

Image Credits
• Format: https://www.flickr.com/photos/zaqography/3835692243/
• Get right data: https://www.flickr.com/photos/encouragement/14759554777/
• Feed: https://www.flickr.com/photos/glutnix/4291194/
• Validate: https://www.flickr.com/photos/lord-jim/16827236591/
• Save: http://www.cnn.com/2015/09/13/living/candice-swanepoel-victorias-secret-model-falls-feat/
• Ship It: https://www.flickr.com/photos/oneeighteen/15492277272/
• Save Predictions: https://www.flickr.com/photos/eelssej_/486414113/
• Get it right: https://www.flickr.com/photos/clickflashphotos/3402287993/
• Update it: https://www.flickr.com/photos/dullhunk/5497202855/
• Review: https://www.flickr.com/photos/pluggedmind/10714537023/

Practical Predictive Modeling in Python

Recommended

More Related Content

What's hot (20)

Viewers also liked (16)

Similar to Practical Predictive Modeling in Python (20)

More from Robert Dempsey (20)

Recently uploaded (20)

Practical Predictive Modeling in Python