Train a new model from scratch

usage: radon-defect-predictor train [-h] [--balancers BALANCERS] [--normalizers NORMALIZERS] path_to_csv classifiers

positional arguments:
  path_to_csv           the path to the csv file containing the data for training
  classifiers           a list of classifiers to train. Possible choices [dt, logit, nb, rf, svm]


optional arguments:
  -h, --help            show this help message and exit
  --balancers BALANCERS
                        a list of balancer to balance training data. Possible choices [none, rus, ros]
  --normalizers NORMALIZERS
                        a list of normalizers to normalize data. Possible choices [none, minmax, std]

path_to_csv

radon-defect-predictor train --path-to-csv path/to/repository-data.csv

The path to the training data (a .csv file). You can generate training data for IaC defect-prediction through radon-miner. An example observation is the following:

filepath	commit	committed_at	failure_prone	metric_1	...	metric_n
roles/tasks/main.yml	25c04...	1526444640	1	value_1	...	value_n

• filepath: string is the path to the file from the repository root;
• commit: string is the commit sha the file belongs to. In per-release based defect-prediction, it is the commit sha of a release, and it is used to group observations of the same release;
• committed_at: string is the commit datetime. In release-based defect-prediction, it is the release date. In just-in-time defect-prediction is the commit date. It is used to sort releases/commits for walk-forward validation;
• failure_prone: integer 1 if the observation is failure-prone; 0 otherwise;
• metric_i: float a metric.

classifiers

radon-defect-predictor train --classifiers="dt logit nb rf svm"

• dt - Train a model using a sklearn.tree.DecisionTreeClassifier classifier;

• logit - Train a model using a sklearn.linear_model.LogisticRegression classifier;

• nb - Train a model using a sklearn.naive_bayes.GaussianNB classifier;

• rf - Train a model using a sklearn.ensemble.RandomForestClassifier classifier;

• svm - Train a model using a sklearn.tree.DecisionTreeClassifier classifier.

--balancers

radon-defect-predictor train --balancers="none rus ros"

• none - Do not balance training data;

• rus - Do balance training data using Random Under-Sampling;

• ros - Do balance training data using Random Over-Sampling.

Not providing any options is the same as passing the option none. However, this option can be passed along the others to train the model by either balancing and not balancing the training set.

--normalizers

radon-defect-predictor train --normalizers="none minmax std"

• none - Do not normalize training data;

• minmax - Transform features by scaling each feature to the range [0,1]. It uses the sklearn.preprocessing.MinMaxScaler;

• std - Standardize features by removing the mean and scaling to unit variance. It uses the sklearn.preprocessing.StandardScaler.

Not providing any options is the same as passing the option none. However, this option can be passed along the others to train the model by either normalizing and not normalizing the training set.

Example

Assuming we are training a new model for the ansible-community/molecule project, download the following training set molecule.csv generated using radon-miner. This is the "ground truth" to train a model for that project.

You can now run the radon-defect-predictor train ... wherever on your system. For the sake of example, let's create and move to a new working directory:

mkdir radon_example
cd radon_example
mv /home/<user>/Downloads/molecule.csv .

Now run:

radon-defect-predictor train molecule.csv "dt" --balancers "none rus" --normalizers "minmax"

or (equivalent)

radon-defect-predictor train molecule.csv "dt" -b "none rus" -n "minmax"

The previous command loads and prepares the .csv file. Then, it builds a model using:

• the Decision Tree classifier ("dt");
• the Random Under-Sampling technique to balance the training data (rus), or none (none);
• the minmax normalization to scale data within the range [0,1] (minmax).

The built model (radondp_model.joblib) is saved into the current working directory. You can see it by running:

ls

molecule.csv
radondp_model.joblib

You can run the same command with different combinations of balancers, normalizers, and classifiers, as explained in previous sections.