def eval_routine(DATA_PATH, TEST_PATH, ITEM_FEATURES_PATH, USER_FEATURES_PATH,
TEST_SIZE_WEEKS, N_POPULAR_ITEMS, INIT_NUM_RECS, N_FIN_RECS):
data_train_lvl_1 = get_raw_data_splits(DATA_PATH, mode=0)
data_val_lvl_1 = get_raw_data_splits(TEST_PATH, mode=0)
item_features, user_features = preprare_features(ITEM_FEATURES_PATH,
USER_FEATURES_PATH)
data_train_lvl_1 = prefilter_items(data_train_lvl_1,
N_POPULAR_ITEMS) # Prefilter routine
data_val_lvl_1 = prefilter_items(data_val_lvl_1,
N_POPULAR_ITEMS) # Prefilter routine
itemid_to_price = get_price_list(data_train_lvl_1, data_val_lvl_1)
user_bought_history = get_bought_ever_list(data_train_lvl_1)
item_to_commodity = get_item_commodities_list(item_features)
recommender = MainRecommender(data_train_lvl_1, itemid_to_price)
result_lvl_1 = data_val_lvl_1.groupby(
'user_id')['item_id'].unique().reset_index()
result_lvl_1.columns = ['user_id', 'actual']
result_lvl_1['base_rec'] = result_lvl_1['user_id'].apply(
lambda x: recommender.get_own_recommendations(x, N=INIT_NUM_RECS))
# result_lvl_1['als_rec'] = result_lvl_1['user_id'].apply(lambda x: recommender.get_als_recommendations(x, N=INIT_NUM_RECS))
result_lvl_1 = postfilter_items(result_lvl_1,
recommender.overall_top_purchases,
item_to_commodity,
itemid_to_price,
user_bought_history,
n=N_FIN_RECS)
res = result_lvl_1.apply(lambda row: money_precision_at_k(
row['result'], row['actual'], itemid_to_price, k=5),
axis=1).mean()
print(f"Result money precision @ 5 metric: {res}")
if SAVE_RESULTS:
result_lvl_1[['user_id', 'result']].to_csv('YN_recs.csv', index=False)
return res
def fit(self, data_train, data_val, item_sub_comm, prices, N=5):
param_grid_list = self._get_param_grid(self.model_params)
result_lvl_1 = data_val.groupby(
'user_id')['item_id'].unique().reset_index()
result_lvl_1.columns = ['user_id', 'actual']
best_params = []
best_score = 0
if self.recommender == 'MainRecommender':
for top_n in self.top_n_list:
X = prefilter_items(data_train, take_n_popular=top_n)
for weighting in self.weighting_list:
model = MainRecommender(X,
weighting=weighting,
**param_grid_list[0])
result_lvl_1['recs'] = result_lvl_1['user_id'].apply(
lambda x: model.get_main_model_recommendations(
x, N=top_n))
result_lvl_1 = postfilter_items(result_lvl_1,
'recs',
item_sub_comm,
prices,
N=N)
score = result_lvl_1.apply(lambda row: self.scoring_func(
row['postfilter_recs'], row['actual'], k=200),
axis=1).mean()
if score > best_score:
best_score = score
best_params = [top_n, weighting, param_grid_list[0]]
for param in param_grid_list[1:]:
model.model = model.fit(model.user_item_matrix,
**param)
result_lvl_1['recs'] = result_lvl_1['user_id'].apply(
lambda x: model.get_main_model_recommendations(
x, N=top_n))
result_lvl_1 = postfilter_items(result_lvl_1,
'recs',
item_sub_comm,
prices,
N=N)
score = result_lvl_1.apply(
lambda row: self.scoring_func(
row['postfilter_recs'], row['actual'], k=N),
axis=1).mean()
if score > best_score:
best_score = score
best_params = [top_n, weighting, param]
return best_score, best_params
def calc_precision_take_n_popular(data_train, data_test, item_features, take_n_popular):
data_train = prefilter_items(data_train, item_features=item_features, take_n_popular=take_n_popular)
als_model = MainRecommender(data_train)
result = data_test.groupby('user_id')['item_id'].unique().reset_index()
result.columns=['user_id', 'actual']
# result['als_recommendations'] = result['user_id'].apply(lambda x: als_model.get_als_recommendations(x, N=5))
result['own_recommendations'] = result['user_id'].apply(lambda x: als_model.get_own_recommendations(x, N=5))
# result['similar_items_recommendations'] = result['user_id'].apply(lambda x: als_model.get_similar_items_recommendation(x, N=5))
res = {}
# res['als_recommendations'] = result.apply(lambda row: precision_at_k(row['als_recommendations'], row['actual']), axis=1).mean()
res['own_recommendations'] = result.apply(lambda row: precision_at_k(row['own_recommendations'], row['actual']), axis=1).mean()
# res['similar_items_recommendations'] = result.apply(lambda row: precision_at_k(row['similar_items_recommendations'], row['actual']), axis=1).mean()
return res, result
def get_results(main_data_path, item_features_path, user_features_path,
val_lvl_1_size_weeks, val_lvl_2_size_weeks, model_lvl_1_path,
model_lvl_2_path, test_data_path):
print('Reading data...')
data = pd.read_csv(main_data_path)
item_features = pd.read_csv(item_features_path)
user_features = pd.read_csv(user_features_path)
item_features.columns = [col.lower() for col in item_features.columns]
user_features.columns = [col.lower() for col in user_features.columns]
item_features.rename(columns={'product_id': 'item_id'}, inplace=True)
user_features.rename(columns={'household_key': 'user_id'}, inplace=True)
data_train_lvl_1 = data[data['week_no'] < data['week_no'].max() -
(val_lvl_1_size_weeks + val_lvl_2_size_weeks)]
data_val_lvl_1 = data[(data['week_no'] >= data['week_no'].max() -
(val_lvl_1_size_weeks + val_lvl_2_size_weeks))
& (data['week_no'] < data['week_no'].max() -
(val_lvl_2_size_weeks))]
data_train_lvl_2 = data_val_lvl_1.copy()
data_val_lvl_2 = data[data['week_no'] >= data['week_no'].max() -
val_lvl_2_size_weeks]
prices = data.groupby('item_id')[['sales_value',
'quantity']].sum().reset_index()
prices['price'] = prices['sales_value'] / prices['quantity']
prices.replace(np.inf, 0, inplace=True)
prices = dict(zip(prices['item_id'], prices['price']))
items_sub_comm = dict(
zip(item_features['item_id'], item_features['sub_commodity_desc']))
trans = DataTransformer()
if test_data_path:
test_data = pd.read_csv(test_data_path)
if model_lvl_1_path and model_lvl_2_path:
print('Reading models...')
with open(model_lvl_1_path, 'rb') as f:
model_lvl_1 = pickle.load(f)
with open(model_lvl_2_path, 'rb') as f:
model_lvl_2 = pickle.load(f)
if test_data_path:
result = get_recommendation_df(test_data, model_lvl_1, model_lvl_2,
items_sub_comm, prices,
item_features, user_features)
else:
result = get_recommendation_df(data_val_lvl_2, model_lvl_1,
model_lvl_2, items_sub_comm, prices,
item_features, user_features)
else:
print('Prepare data for level 1 model fit...')
data_train_lvl_1 = prefilter_items(data_train_lvl_1, 6000)
print('Level 1 model fit...')
num_threads = multiprocessing.cpu_count() // 2 + 1
model_lvl_1 = MainRecommender(data_train_lvl_1,
weighting=None,
n_factors=100,
regularization=0.01,
iterations=100,
num_threads=num_threads)
print('Construct level 1 recommendations DataFrame...')
result_lvl_1 = get_recommendation_lvl_1(data_val_lvl_1, model_lvl_1)
result_lvl_1 = postfilter_items(result_lvl_1,
'lvl_1_recs',
items_sub_comm,
prices,
N=200)
result_lvl_1.rename(
columns={'postfilter_lvl_1_recs': 'recommendations'}, inplace=True)
print('Prepare data for level 2 model fit...')
data_train_lvl_2 = trans.fit_transform(result_lvl_1,
data_val_lvl_1,
item_features,
user_features,
with_targets=True)
y = data_train_lvl_2['target']
X = data_train_lvl_2.drop('target', axis=1)
cat_features = [
'department', 'brand', 'commodity_desc', 'sub_commodity_desc',
'curr_size_of_product', 'age_desc', 'marital_status_code',
'income_desc', 'homeowner_desc', 'hh_comp_desc',
'household_size_desc', 'kid_category_desc'
]
class_1_weight = len(y[y == 0]) / len(y[y == 1])
model_lvl_2 = CatBoostClassifier(n_estimators=300,
max_depth=7,
class_weights=[1, class_1_weight],
cat_features=cat_features)
print('Level 2 model fit...')
model_lvl_2.fit(X, y)
if test_data_path:
result = get_recommendation_df(test_data, model_lvl_1, model_lvl_2,
items_sub_comm, prices,
item_features, user_features)
else:
result = get_recommendation_df(data_val_lvl_2, model_lvl_1,
model_lvl_2, items_sub_comm, prices,
item_features, user_features)
print('Calculating final metric...')
money_precision_at_5 = result.apply(
lambda row: money_precision_at_k(row['postfilter_catboost_recs'], row[
'actual'], row['recommend_prices']),
axis=1).mean()
print(
f'Money precision@5 for final recommendations:{money_precision_at_5}')
result.to_csv('Final_recommendations.csv')
return result
from src.utils import prefilter_items, postfilter, create_dataset, dataset_processing
from src.recommenders import MainRecommender, Ranking
data = pd.read_csv('../raw_data/retail_train.csv')
item_features = pd.read_csv('../raw_data/product.csv')
user_features = pd.read_csv('../raw_data/hh_demographic.csv')
# column processing
item_features.columns = [col.lower() for col in item_features.columns]
user_features.columns = [col.lower() for col in user_features.columns]
item_features.rename(columns={'product_id': 'item_id'}, inplace=True)
user_features.rename(columns={'household_key': 'user_id'}, inplace=True)
# data filtering
data = prefilter_items(data, item_features=item_features, take_n_popular=5000)
candidates = pd.DataFrame(data['user_id'].unique())
candidates = candidates.rename(columns={0: 'user_id'})
recommender = MainRecommender(data)
# Рекомендации по BM25 взвешиванию
candidates['candidates'] = candidates['user_id'].apply(lambda x: recommender.get_bm25_recommendations(x, N=100))
# Создадим датафрейм в целевой переменной и добавим фичи юзеров и товаров
targets = create_dataset(data=data, data_candidates=candidates, users_info=user_features, items_info=item_features)
# Сгенерируем новые фичи
targets = dataset_processing(dataset=targets, data=data, items_info=item_features)
def __init__(self,
data,
item_info,
weighting=True,
first_model_weeks=6,
second_model_weeks=3,
take_n_popular=7000):
self.top_purchases = data.groupby(
['user_id', 'item_id'])['quantity'].count().reset_index()
self.top_purchases.sort_values('quantity',
ascending=False,
inplace=True)
self.top_purchases = self.top_purchases[
self.top_purchases['item_id'] != 999999]
self.overall_top_purchases = data.groupby(
'item_id')['quantity'].count().reset_index()
self.overall_top_purchases.sort_values('quantity',
ascending=False,
inplace=True)
self.overall_top_purchases = self.overall_top_purchases[
self.overall_top_purchases['item_id'] != 999999]
self.overall_top_purchases = self.overall_top_purchases.item_id.tolist(
)
self.user_buyses = data.groupby(
'user_id')['item_id'].unique().reset_index()
self.user_buyses.columns = ['user_id', 'actual']
self.user_item_matrix = self._prepare_matrix(data)
self.id_to_itemid, self.id_to_userid, \
self.itemid_to_id, self.userid_to_id = self._prepare_dicts(self.user_item_matrix)
if weighting:
self.user_item_matrix = bm25_weight(self.user_item_matrix.T).T
self.model = self.fit(self.user_item_matrix)
self.own_recommender = self.fit_own_recommender(self.user_item_matrix)
self.item_info = item_info
self.first_model_weeks = first_model_weeks
self.second_model_weeks = second_model_weeks
self.val_lvl_1_size_weeks = first_model_weeks
self.val_lvl_2_size_weeks = second_model_weeks
self.data_train_lvl_1 = data[
data['week_no'] < data['week_no'].max() -
(self.val_lvl_1_size_weeks + self.val_lvl_2_size_weeks)]
self.data_val_lvl_1 = data[
(data['week_no'] >= data['week_no'].max() -
(self.val_lvl_1_size_weeks + self.val_lvl_2_size_weeks))
& (data['week_no'] < data['week_no'].max() -
(self.val_lvl_2_size_weeks))]
self.data_train_lvl_1 = prefilter_items(self.data_train_lvl_1,
item_features=self.item_info,
take_n_popular=take_n_popular)
self.data_train_lvl_2 = self.data_val_lvl_1.copy()
self.data_val_lvl_2 = data[data['week_no'] >= data['week_no'].max() -
self.val_lvl_2_size_weeks]
self.user_buyses_lvl_1 = self.data_train_lvl_1.groupby(
'user_id')['item_id'].unique().reset_index()
self.user_buyses_lvl_1.columns = ['user_id', 'actual']
self.users_recommendations_lvl_1 = self.get_als_recommendations_users(
self.user_buyses_lvl_1["user_id"], N=25)