features.py 文件源码

def get_numerical_features(fnum, fname, df, nvalues, dt,
                           sentinel, logt, plevel):
    r"""Transform numerical features with imputation and possibly
    log-transformation.

    Parameters
    ----------
    fnum : int
        Feature number, strictly for logging purposes
    fname : str
        Name of the numerical column in the dataframe ``df``.
    df : pandas.DataFrame
        Dataframe containing the column ``fname``.
    nvalues : int
        The number of unique values.
    dt : str
        The values ``'float64'``, ``'int64'``, or ``'bool'``.
    sentinel : float
        The number to be imputed for NaN values.
    logt : bool
        If ``True``, then log-transform numerical values.
    plevel : float
        The p-value threshold to test if a feature is normally distributed.

    Returns
    -------
    new_values : numpy array
        The set of imputed and transformed features.

    """
    feature = df[fname]
    if len(feature) == nvalues:
        logger.info("Feature %d: %s is a numerical feature of type %s with maximum number of values %d",
                    fnum, fname, dt, nvalues)
    else:
        logger.info("Feature %d: %s is a numerical feature of type %s with %d unique values",
                    fnum, fname, dt, nvalues)
    # imputer for float, integer, or boolean data types
    new_values = impute_values(feature, dt, sentinel)
    # log-transform any values that do not fit a normal distribution
    if logt and np.all(new_values > 0):
        stat, pvalue = sps.normaltest(new_values)
        if pvalue <= plevel:
            logger.info("Feature %d: %s is not normally distributed [p-value: %f]",
                        fnum, fname, pvalue)
            new_values = np.log(new_values)
    return new_values


#
# Function get_polynomials
#