def make_order_interval_features(orders: pd.DataFrame) -> pd.DataFrame:
orders = orders.sort_values(['client_id', 'datetime'])
last_order_client = orders['client_id'].shift(1)
is_same_client = last_order_client == orders['client_id']
orders['last_order_datetime'] = orders['datetime'].shift(1)
orders['orders_interval'] = np.nan
orders.loc[is_same_client, 'orders_interval'] = (
orders.loc[is_same_client, 'datetime'] -
orders.loc[is_same_client, 'last_order_datetime']
).dt.total_seconds() / SECONDS_IN_DAY
cl_gb = orders.groupby('client_id', sort=False)
features = cl_gb.agg(
{
'orders_interval': [
'mean', # mean interval between orders
'median',
'std', # constancy of orders
'min',
'max',
'last', # interval between last 2 orders
]
}
)
drop_column_multi_index_inplace(features)
features.reset_index(inplace=True)
features.fillna(-3, inplace=True)
return features
def make_store_features(orders: pd.DataFrame) -> pd.DataFrame:
cl_st_gb = orders.groupby(['client_id', 'store_id'])
store_agg = cl_st_gb.agg({
'transaction_id': ['count'],
})
drop_column_multi_index_inplace(store_agg)
store_agg.reset_index(inplace=True)
cl_gb = store_agg.groupby(['client_id'])
simple_features = cl_gb.agg(
{
'transaction_id_count': ['max', 'mean', 'median']
}
)
drop_column_multi_index_inplace(simple_features)
simple_features.reset_index(inplace=True)
simple_features.columns = (
['client_id'] +
[
f'store_{col}'
for col in simple_features.columns[1:]
]
)
latent_features = make_latent_store_features(orders)
features = pd.merge(
simple_features,
latent_features,
on='client_id'
)
return features
def make_small_product_features(purchases: pd.DataFrame) -> pd.DataFrame:
cl_pr_gb = purchases.groupby(['client_id', 'product_id'])
product_agg = cl_pr_gb.agg({
'product_quantity': ['sum'],
})
drop_column_multi_index_inplace(product_agg)
product_agg.reset_index(inplace=True)
cl_gb = product_agg.groupby(['client_id'])
features = cl_gb.agg({'product_quantity_sum': ['max']})
drop_column_multi_index_inplace(features)
features.reset_index(inplace=True)
return features
def make_features_for_orders_with_express_points_spent(
orders: pd.DataFrame
) -> pd.DataFrame:
orders_with_eps = orders.loc[orders['express_points_spent'] != 0]
o_gb = orders_with_eps.groupby(['client_id'])
features = o_gb.agg(
{
'purchase_sum': ['median'],
'datetime': ['max']
}
)
drop_column_multi_index_inplace(features)
features.reset_index(inplace=True)
features['days_from_last_express_points_spent'] = (
MAILING_DATETIME - features['datetime_max']
).dt.days
features.drop(columns=['datetime_max'], inplace=True)
features.rename(
columns={
'purchase_sum_median': 'median_purchase_sum_eps'
},
inplace=True,
)
order_int_features = make_order_interval_features(orders_with_eps)
renamings = {
col: f'{col}_eps'
for col in order_int_features
if col != 'client_id'
}
order_int_features.rename(columns=renamings, inplace=True)
features = pd.merge(
features,
order_int_features,
on='client_id',
)
features = features.merge(
pd.Series(orders['client_id'].unique(), name='client_id'),
how='right',
)
return features
def make_really_purchase_features(purchases: pd.DataFrame) -> pd.DataFrame:
simple_purchases = purchases.reindex(
columns=['client_id', 'product_id', 'trn_sum_from_iss']
)
prices_bounds = [0, 98, 195, 490, 950, 1900, 4400, FLOAT32_MAX]
agg_dict = {}
for i, lower_bound in enumerate(prices_bounds[:-1]):
upper_bound = prices_bounds[i + 1]
name = f'price_from_{lower_bound}'
simple_purchases[name] = (
(simple_purchases['trn_sum_from_iss'] >= lower_bound) &
(simple_purchases['trn_sum_from_iss'] < upper_bound)
).astype(int)
agg_dict[name] = ['sum', 'mean']
agg_dict.update(
{
'trn_sum_from_iss': ['median'], # median product price
'product_id': ['count', 'nunique'],
}
)
simple_features = simple_purchases.groupby('client_id').agg(agg_dict)
drop_column_multi_index_inplace(simple_features)
simple_features.reset_index(inplace=True)
p_gb = purchases.groupby(['client_id', 'transaction_id'])
purchase_agg = p_gb.agg(
{
'product_id': ['count'],
'product_quantity': ['max'],
}
)
drop_column_multi_index_inplace(purchase_agg)
purchase_agg.reset_index(inplace=True)
o_gb = purchase_agg.groupby('client_id')
complex_features = o_gb.agg(
{
# mean products in order
'product_id_count': ['mean', 'median'],
# mean max number of one product
'product_quantity_max': ['mean', 'median'],
}
)
drop_column_multi_index_inplace(complex_features)
complex_features.reset_index(inplace=True)
features = pd.merge(
simple_features,
complex_features,
on='client_id'
)
return features
def make_order_features(orders: pd.DataFrame) -> pd.DataFrame:
orders = orders.copy()
o_gb = orders.groupby('client_id')
agg_dict = {
'transaction_id': ['count'], # number of orders
'regular_points_received': ['sum', 'max', 'median'],
'express_points_received': ['sum', 'max', 'median'],
'regular_points_spent': ['sum', 'min', 'median'],
'express_points_spent': ['sum', 'min', 'median'],
'purchase_sum': ['sum', 'max', 'median'],
'store_id': ['nunique'], # number of unique stores
'datetime': ['max'], # datetime of last order
}
# is regular/express points spent/received
for points_type in POINT_TYPES:
for event_type in POINT_EVENT_TYPES:
col_name = f'{points_type}_points_{event_type}'
new_col_name = f'is_{points_type}_points_{event_type}'
orders[new_col_name] = (orders[col_name] != 0).astype(int)
agg_dict[new_col_name] = ['sum']
features = o_gb.agg(agg_dict)
drop_column_multi_index_inplace(features)
features.reset_index(inplace=True)
features['days_from_last_order'] = (
MAILING_DATETIME - features['datetime_max']
).dt.total_seconds() // SECONDS_IN_DAY
features.drop(columns=['datetime_max'], inplace=True)
# proportion of regular/express points spent to all transactions
for points_type in POINT_TYPES:
for event_type in POINT_EVENT_TYPES:
col_name = f'is_{points_type}_points_{event_type}_sum'
new_col_name = f'proportion_count_{points_type}_points_{event_type}'
features[new_col_name] = (
features[col_name] / features['transaction_id_count']
)
express_col = f'is_express_points_spent_sum'
regular_col = f'is_regular_points_spent_sum'
new_col_name = f'ratio_count_express_to_regular_points_spent'
features[new_col_name] = (
features[express_col] / features[regular_col]
).replace(np.inf, FLOAT32_MAX)
for points_type in POINT_TYPES:
spent_col = f'is_{points_type}_points_spent_sum'
received_col = f'is_{points_type}_points_received_sum'
new_col_name = f'ratio_count_{points_type}_points_spent_to_received'
features[new_col_name] = (
features[spent_col] / features[received_col]
).replace(np.inf, 1000)
for points_type in POINT_TYPES:
spent_col = f'{points_type}_points_spent_sum'
orders_sum_col = f'purchase_sum_sum'
new_col_name = f'ratio_sum_{points_type}_points_spent_to_purchases_sum'
features[new_col_name] = features[spent_col] / features[orders_sum_col]
new_col_name = f'ratio_sum_express_points_spent_to_sum_regular_points_spent'
regular_col = f'regular_points_spent_sum'
express_col = f'express_points_spent_sum'
features[new_col_name] = features[express_col] / features[regular_col]
return features