Dataset generation

The smote_cd package comes with functions to generate synthetic datasets having compositional labels. The dataset is created using a Dirichlet distribution of parameter \(\alpha = X \beta\), \(X\) being the features of the dataset and \(\beta\) the matrix of the regression coefficients.

Such a matrix \(\beta\) can be created with the function smote_cd.dataset_generation.generate_betas():

betas = smote_cd.dataset_generation.generate_betas(n_features,n_classes)

The shape of the matrix \(\beta\) is (n_classes, n_features+1) because the first column is associated to the intercept. Then, knowing such a matrix, one can generate a synthetic dataset of size size:

X, y, _ = smote_cd.dataset_generation.generate_dataset(n_features, n_classes, size, betas=betas)

The third argument returned by the function is actually the \(\beta\) matrix. Indeed, if no parameter betas is given, the function will create a new matrix \(\beta\). Of course, if you want reproducible results or to generate several datasets with the same distribution, it is better to first generate your matrix and then give it as an input of the function.

Another way to keep the same betas is to get the one returned by the function to reuse it:

X_1, y_1, betas = smote_cd.dataset_generation.generate_dataset(n_features, n_classes, size_1)
X_2, y_2, _ = smote_cd.dataset_generation.generate_dataset(n_features, n_classes, size_2, betas=betas)

Note that the parameter random_state will set the random seed to its given value. If no betas is specified, its creation will also be fixed by the random seed. It means that the two following calls return the same result:

X_1, y_1, _ = smote_cd.dataset_generation.generate_dataset(n_features, n_classes, size, random_seed=0)
X_2, y_2, _ = smote_cd.dataset_generation.generate_dataset(n_features, n_classes, size, random_seed=0)

But these two do not return the same result:

X_1, y_1, _ = smote_cd.dataset_generation.generate_dataset(n_features, n_classes, size, betas=betas_1, random_seed=0)
X_2, y_2, _ = smote_cd.dataset_generation.generate_dataset(n_features, n_classes, size, betas=betas_2, random_seed=0)

More example are available on the page smote_cd.dataset_generation.generate_dataset().