for day in tqdm(range(18, 26)):
count = temp.groupby([feat_1]).apply(lambda x: x['is_trade'][(x['context_date_day'] < day).values].count()).\
reset_index(name=feat_1 + '_all')
count1 = temp.groupby([feat_1]).apply(lambda x: x['is_trade'][(x['context_date_day'] < day).values].sum()).\
reset_index(name=feat_1 + '_1')
count[feat_1 + '_1'] = count1[feat_1 + '_1']
# TODO: should handle first day conversion count and sum ?
count.fillna(value=0, inplace=True)
count['context_date_day'] = day
res = res.append(count, ignore_index=True)
# only smooth item_id and item_brand_id here
if feat_1 == 'item_id':
print('smoothing item_id')
bs_item = BayesianSmoothing(1, 1)
bs_item.update(res[feat_1 + '_all'].values, res[feat_1 + '_1'].values,
1000, 0.001)
res[feat_1 + '_smooth'] = (res[feat_1 + '_1'] + bs_item.alpha) / \
(res[feat_1 + '_all'] + bs_item.alpha + bs_item.beta)
if feat_1 == 'item_brand_id':
print('smoothing item_brand_id')
bs_brand = BayesianSmoothing(1, 1)
bs_brand.update(res[feat_1 + '_all'].values, res[feat_1 + '_1'].values,
1000, 0.001)
res[feat_1 + '_smooth'] = (res[feat_1 + '_1'] + bs_brand.alpha) / \
(res[feat_1 + '_all'] + bs_brand.alpha + bs_brand.beta)
# all features conversion rate
res[feat_1 + '_rate'] = res[feat_1 + '_1'] / res[feat_1 + '_all']
train = train.merge(res, how='left', on=[feat_1, 'context_date_day'])
for day in tqdm(range(18, 26)):
count = temp.groupby([feat_1]).apply(lambda x: x['is_trade'][(x['context_date_day'] < day).values].count()).\
reset_index(name=feat_1 + '_all')
count1 = temp.groupby([feat_1]).apply(lambda x: x['is_trade'][(x['context_date_day'] < day).values].sum()).\
reset_index(name=feat_1 + '_1')
count[feat_1 + '_1'] = count1[feat_1 + '_1']
# TODO: should handle first day conversion count and sum ?
count.fillna(value=0, inplace=True)
count['context_date_day'] = day
res = res.append(count, ignore_index=True)
# only smooth shop_id here
if feat_1 == 'shop_id':
print('smoothing shop_id')
bs = BayesianSmoothing(1, 1)
bs.update(res[feat_1 + '_all'].values, res[feat_1 + '_1'].values, 1000,
0.001)
res[feat_1 + '_smooth'] = (res[feat_1 + '_1'] + bs.alpha) / (
res[feat_1 + '_all'] + bs.alpha + bs.beta)
# all features conversion rate
res[feat_1 + '_rate'] = res[feat_1 + '_1'] / res[feat_1 + '_all']
train = train.merge(res, how='left', on=[feat_1, 'context_date_day'])
test = test.merge(res, how='left', on=[feat_1, 'context_date_day'])
if feat_1 == 'shop_id':
train['shop_id_smooth'] = train['shop_id_smooth'].fillna(
value=bs.alpha / (bs.alpha + bs.beta))
test['shop_id_smooth'] = test['shop_id_smooth'].fillna(
value=bs.alpha / (bs.alpha + bs.beta))
for day in tqdm(range(18, 26)):
count = temp.groupby(['item_category_1']).apply(lambda x: x['is_trade'][(x['context_date_day'] < day).values].count()).\
reset_index(name='item_category_1' + '_all')
count1 = temp.groupby(['item_category_1']).apply(lambda x: x['is_trade'][(x['context_date_day'] < day).values].sum()).\
reset_index(name='item_category_1' + '_1')
count['item_category_1' + '_1'] = count1['item_category_1' + '_1']
# TODO: should handle first day conversion count and sum ?
count.fillna(value=0, inplace=True)
count['context_date_day'] = day
res = res.append(count, ignore_index=True)
print('smoothing category_id')
bs = BayesianSmoothing(1, 1)
bs.update(res['item_category_1' + '_all'].values,
res['item_category_1' + '_1'].values, 1000, 0.001)
res['item_category_1' + '_smooth'] = (res['item_category_1' + '_1'] + bs.alpha) / \
(res['item_category_1' + '_all'] + bs.alpha + bs.beta)
print('item_category_1', bs.alpha, bs.beta)
# item_category_1 2.16580301337 66.9451051993
# conversion rate
res['item_category_1' +
'_rate'] = res['item_category_1' + '_1'] / res['item_category_1' + '_all']
res.to_pickle('../features/concat_cate_smt_ctr_feature_304.p')
# ================================================
# property part
# ================================================