Preprocessing Code: spam_outlier_detection.py

Requires:

• numpy
• pandas
• sklearn
• matplotlib

Usage:

Option to produce .data files without outliers:

Run the code with 2 as a commandline argument:

python spam_outlier_detection.py 2

Output: spambase_iso.data, spambase_lof.data

Option to produce tSNE graphs

Run the code with {0,1} {0,1} {2,3} as the commandline arguments:

python spam_outlier_detection.py (1) (2) (3)

(1): Options for the outlier method:

• 0: Isolation Forest
• 1: Local Outlier Factor

(2): Options for graph content:

• 0: graph spam vs non-spam
• 1: graph outliers vs non-outliers

(3): Options for the tSNE dimension:

• 2: two-dimensional tSNE graph
• 3: tree-dimensional tSNE graph

Output: 2-d or 3-d tSNE graph.

Example:

python spam_outlier_detection.py 0 0 2

Utilizing Isolation Forest it will graph spam vs not-spam on 2 tSNE dimensions.

ANN Code: ann.py

Requires:

• keras
• tensorflow
• pandas
• sklearn
• matplotlib

Set up:

To perform a grid search over the number of nodes per layer and number of hidden layers, change max_hidden_layers in main() to one of the options described below. Set the learning rate by changing the lr variable in main and set the value of k for k-fold validation by changing k (again, options below). Setting the activation function, output layer activation function, and loss function is achieved by changing activation_fn, output_fn, and loss_fn respectively. Comment out the call to testing() at the end of the file and call main(). At the end of each fold, the training and testing misclassification error matrices will be printed, where the row corresponds to the number of hidden layers and the column corresponds to the number of nodes per layer.

To generate the generalized error, ROC, and PR plots, comment out the call to main() at the end of the file and call testing(). In testing(), set the number of hidden layers and nodes per layer by changing hidden_layers and nodes_per_layer, change the learning rate by setting lr, and set the activation, output and loss functions by changing activation_fn, output_fn, and loss_fn.

Usage: python3 ann.py

Options for activation/output function:

• "sigmoid"
• "linear"
• layers.LeakyReLU(alpha=0.01)
• "tanh"

Options for number of hidden layers:

• Any nonzero positive integer

Options for number of nodes per layer:

• Any nonzero positive integer (or list of positive integers if setting in main)

Options for learning rate:

• Any nonzero positive number

Options for loss function:

• "binary cross-entropy"
• "mse"

Bayes Code: bayes.py

Requires:

• numpy

Usage:

python bayes.py filepath method transformation

Options for method:

• 0: assume gaussian distribution
• 1: discrete variable calculation

Options for transformation:

• <=0: no transformation of data
• 1: transform data using a mean-bucketed method
• >=2: transform data using specified number of quantiles

ex:

"python bayes.py spambase.data 1 3" will use 3 buckets and calculate likelihood using a discrete function

Output:

Prints TP/FP/TN/FN counts and misclassification for each batch of training and testing over 10 folds Also prints the average training and testing misclassification at the end of training as well as min testing misclassification.

Supplemental: bayes_classifier_extra.py

Slightly different classification style created as a reference, and on different but similar data, to help contextualize the performance of/refine the methodology our main classifier. Not involved in the Spambase comparison.

Bayes graphing: bayes_graphing.py

Requires:

• matplotlib

Usages:

python bayes_graphing.py filepath