def my_eval(pre_y, valid_df, answer, phase=-1, print_mark=True):
# 构造submit csv
valid_submit = utils.save_pre_as_submit_format_csv(valid_df, pre_y)
submit_csv_path = utils.save(valid_submit,
file_dir='./cache/tmp_phase_submit')
# 构造truth csv
valid_answer = valid_df.loc[:, ['user_id']].drop_duplicates(['user_id'],
keep='first')
# answer中user列唯一、item列也是唯一
valid_answer = valid_answer.merge(answer, on='user_id', how='left')
valid_answer_save_path = './cache/tmp_phase_submit/valid_answer.csv'
valid_answer = valid_answer[['phase_id', 'user_id', 'item_id', 'item_deg']]
valid_answer.to_csv(valid_answer_save_path, index=False, header=False)
score, \
ndcg_50_full, ndcg_50_half, \
hitrate_50_full, hitrate_50_half = evaluate(submit_csv_path, valid_answer_save_path,
recall_num=None)
if print_mark:
print(
'phase:{}, score:{}, ndcg_50_full:{}, ndcg_50_half:{}, hitrate_50_full:{}, hitrate_50_half:{}'
.format(phase, score, ndcg_50_full, ndcg_50_half, hitrate_50_full,
hitrate_50_half))
return np.array(
[score, ndcg_50_full, ndcg_50_half, hitrate_50_full,
hitrate_50_half]).reshape(-1, )
'item_id': np.str
})
if conf.subsampling:
recall_feature_df = recall_feature_df[
recall_feature_df['user_id'].isin(
one_phase_recall_item_df['user_id'])]
print('load recall features: phase:{} shape:{}'.format(
phase, recall_feature_df.shape[0]))
else:
# featuring
recall_feature_df = do_featuring(
all_phase_click_no_qtime,
recall_sample_df,
hot_df,
conf.process_num,
item_txt_embedding_dim,
is_recall=True,
feature_caching_path=conf.recall_feature_path.format(phase))
submit_x = recall_feature_df[recall_feature_df.columns.difference(
['user_id', 'item_id', 'label'])].values
submit_pre_y = model.predict_proba(submit_x)[:, 1]
submit = utils.save_pre_as_submit_format_csv(recall_sample_df,
submit_pre_y)
submit_all = submit_all.append(submit)
print('--------------------------- 保存预测文件 --------------------------')
utils.save(submit_all, 50)
# todo 统计不同阶段时间分段情况
# todo user点击深度超过10的情况怎么处理
print('------------------------ 模型训练 start time:{}'.format(time_str))
# submit = train_model_lgb(feature_all, recall_rate=hit_rate, hot_list=hot_list, valid=0.2, topk=50, num_boost_round=1, early_stopping_rounds=1)
# submit = train_model_rf(train_test, recall_rate=1, hot_list=hot_list, valid=0.2, topk=50)
model = rank_rf(train_x, train_y)
# model = rank_xgb(train_x, train_y)
# joblib.dump(model, './cache/rf.pkl')
print('------------------------ 模型验证 ----------------------------')
time_str = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
print(
'------------------------ 模型验证 part1:采样集合验证 start time:{}'.format(
time_str))
pre_y = model.predict_proba(valid_x)[:, 1]
# 构造submit csv
valid_submit = save_pre_as_submit_format_csv(valid_data, pre_y)
submit_csv_path = utils.save(valid_submit, file_dir=temp_result_path)
# 构造truth csv
valid_answer = valid_data.loc[:, ['user_id']].drop_duplicates(
['user_id'], keep='first')
# answer中user列唯一、item列也是唯一
valid_answer = valid_answer.merge(debias_track_answer,
on='user_id',
how='left')
valid_answer_save_path = './cache/tmp_phase_submit/valid_answer.csv'
valid_answer = valid_answer[[
'phase_id', 'user_id', 'item_id', 'item_deg'
]]
valid_answer.to_csv(valid_answer_save_path, index=False, header=False)
score, \