HEPMASS-IMB

HEPMASS-IMB is a benchmark dataset for signal-background classification in High-Energy Physics (HEP), derived from HEPMASS (Baldi et al.) by imbalancing it two times: on the class labels, as well as on the mass labels.

• It has 27 feature columns (named from f0 to f26), and a 28-th mass feature (named mass).
• The 27 features are already normalized to have approximately zero-mean and unitary variance.
• The mass feature has five unique values: 500, 750, 1000, 1250, and 1500.
• There are two class labels: 1 (signal), and 0 (background).
• The dataset describes the decay of an hypothetical particle: \(X \to t\bar{t}\to X->t\bar{t} \to W^+bW^-\bar{b}\).

Further details about the original dataset are available here, whereas a description of our modifications is presented in our paper.

NOTE:

• The files provided here represent only the training-set, since it's what is diverse compared to the original HEPMASS.
• The label column has been renamed from "# label" to "type".
• There are two new columns: name, and weight.

Steps to adapt `all_test.csv` (from HEPMASS):

# 1. Load csv
df = pd.read_csv('<your-path>/all_test.csv')

# 2. Rename columns
df.rename(columns={'# label': 'type'}, inplace=True)

# 3. Adjust mass column
mass = np.sort(df['mass'].unique())
df.loc[df['mass'] == mass[0], 'mass'] = 500.0

# 4. Finally save the new csv
df.to_csv('<your-path>/test.csv', index=False)