def transform(self, X, y=None):
enc = lambda i:str(i) if self.ascategory else i
r = X
for c in date_cols(r):
logger.debug('encoding date column {}'.format(c))
r[c + '_year'] = r[c].apply(lambda ts:enc(ts.year))
r[c + '_month'] = r[c].apply(lambda ts:enc(ts.month))
r[c + '_week'] = r[c].apply(lambda ts:enc(ts.week))
r[c + '_wom'] = r[c].apply(lambda ts:enc((ts.day-1) // 7 + 1))
r[c + '_day'] = r[c].apply(lambda ts:enc(ts.day))
r[c + '_dow'] = r[c].apply(lambda ts:enc(ts.dayofweek))
r[c + '_hour'] = r[c].apply(lambda ts:enc(ts.hour))
r[c] = r[c].apply(lambda ts:float(ts.toordinal() - self.t0.toordinal()))
return r
def _compute_model_correls(df, model='linear', columns=None, sparse=False):
for c in date_cols(df):
df[c] = _index_encode(df[c])
res = pd.DataFrame(index=df.columns, columns=columns)
for i in df.columns:
xi = pd.get_dummies(df.loc[:, i], sparse=sparse)
for j in columns:
if j == i:
res.loc[i, j] = 1.
else:
xj = df.loc[:, j]
rf = _get_model(model, xj)
rf.fit(xi, xj)
res.loc[i, j] = _get_score(model, rf, xi, xj)
del(rf)
return res.astype(float) # not sure why this is needed
def _compute_pearson_correls(df, columns):
for c in object_cols(df) + date_cols(df):
df[c] = _index_encode(df[c])
cor = np.corrcoef(df, rowvar=0)
res = pd.DataFrame(cor, index=df.columns, columns=df.columns)
return res[columns]