最终排名：第二名
融合模型最高分：A榜：0.74062；B榜：0.74347
单模型最高分：A榜：0.73989；B榜：0.7422
赛题2：基于收支记录判断借贷意愿
手机APP的首页首屏，是流量最大的页面，对应页面资源极其珍贵。现要将借钱卡片投放到首页，展现给最有点击意愿的用户。请基于用户真实的历史收支数据来预测用户是否会点击借钱卡片。
features227 = train227[['id']]
用户基础属性227 = pd.read_csv(features_dir227 + "用户基础属性.csv")
features227 = pd.merge(features227, 用户基础属性227, on='id', how='left')
收入总和227 = pd.read_csv(features_dir227 + "收入总和.csv")
features227 = pd.merge(features227, 收入总和227, on='id', how='left')
支出总和227 = pd.read_csv(features_dir227 + "支出总和.csv")
features227 = pd.merge(features227, 支出总和227, on='id', how='left')
收支总和227 = pd.read_csv(features_dir227 + "收支总和.csv")
features227 = pd.merge(features227, 收支总和227, on='id', how='left')
收支记录次数227 = pd.read_csv(features_dir227 + "收支记录次数.csv")
features227 = pd.merge(features227, 收支记录次数227, on='id', how='left')
收入最大值227 = pd.read_csv(features_dir227 + "收入最大值.csv")
features227 = pd.merge(features227, 收入最大值227, on='id', how='left')
支出最大值227 = pd.read_csv(features_dir227 + "支出最大值.csv")
features227 = pd.merge(features227, 支出最大值227, on='id', how='left')
各类别收支记录次数227 = pd.read_csv(features_dir227 + "各类别收支记录次数.csv")
features227 = pd.merge(features227, 各类别收支记录次数227, on='id', how='left')
各类别收支金额227 = pd.read_csv(features_dir227 + "各类别收支金额.csv")
features227 = pd.merge(features227, 各类别收支金额227, on='id', how='left')
# 加入以下两个特征后，有明显提升
各原始类别收支记录次数227 = pd.read_csv(features_dir227 + "各原始类别收支记录次数.csv")
features227 = pd.merge(features227, 各原始类别收支记录次数227, on='id', how='left')
各原始类别收支金额227 = pd.read_csv(features_dir227 + "各原始类别收支金额.csv")
features227 = pd.merge(features227, 各原始类别收支金额227, on='id', how='left')
收入均值227 = pd.read_csv(features_dir227 + "收入均值.csv")
features227 = pd.merge(features227, 收入均值227, on='id', how='left')
支出均值227 = pd.read_csv(features_dir227 + "支出均值.csv")
features227 = pd.merge(features227, 支出均值227, on='id', how='left')
收支均值227 = pd.read_csv(features_dir227 + "收支均值.csv")
features227 = pd.merge(features227, 收支均值227, on='id', how='left')
# 选24个得分最高的原始类别特征，两两相加，得到288维特征，然后再筛选，略有提升
原始类别次数两两相加组合特征227 = pd.read_csv(features_dir227 + "原始类别次数两两相加组合特征.csv")
features227 = pd.merge(features227, 原始类别次数两两相加组合特征227, on='id', how='left')
原始类别金额两两相加组合特征227 = pd.read_csv(features_dir227 + "原始类别金额两两相加组合特征.csv")
features227 = pd.merge(features227, 原始类别金额两两相加组合特征227, on='id', how='left')def pipeline(iteration,random_seed,max_depth,lambd,subsample,colsample_bytree,min_child_weight):
if max_depth==3:
n_estimators = 400
elif max_depth==4:
n_estimators = 300
kf = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
preds306 = np.zeros((X306_csr.shape[0]))
i = 0
models = [0]*kf.get_n_splits()
roc_scores = []
for train_index, val_index in kf.split(X227_csr, Y227):
X_train, X_val = X227_csr[train_index], X227_csr[val_index]
y_train, y_val = Y227[train_index], Y227[val_index]
eval_set = [(X_train, y_train), (X_val, y_val)]
models[i] = xgb.XGBClassifier(n_estimators=n_estimators,
max_depth = max_depth,
reg_lambda = lambd,
subsample = subsample,
colsample_bytree = colsample_bytree,
min_child_weight = min_child_weight,
n_jobs = 6,
random_state=random_seed)
models[i].fit(X_train, y_train, eval_metric=['auc', "error", "logloss"], eval_set=eval_set, verbose=False)
score = models[i].predict_proba(X_val)[:, 1]
roc = roc_auc_score(y_val, score)
roc_scores.append(roc)
print(roc)
preds306 = preds306 + models[i].predict_proba(X306_csr)[:,1]
i+=1
preds306 = preds306 / kf.get_n_splits()
pickle.dump(preds306, open(inter_dir + 'preds306_'+str(iteration)+'.p', 'wb'))
pickle.dump(models, open(inter_dir + 'models_'+str(iteration)+'.p','wb'))
pickle.dump(roc_scores, open(inter_dir + 'roc_scores_'+str(iteration)+'.p','wb'))
print("mean_roc_score: {}".format(np.mean(roc_scores)))
随机参数，训练出30个模型。
random.seed(42)
random_seed = list(range(43))
max_depth = [3,4]
lambd = list(range(0,5))
subsample = [i/1000.0 for i in range(700,800)]
colsample_bytree = [i/1000.0 for i in range(700,800)]
min_child_weight = [i/100.0 for i in range(150,250)]
# n_feature = range(150,282,2)
random.shuffle(random_seed)
random.shuffle(max_depth)
random.shuffle(lambd)
random.shuffle(subsample)
random.shuffle(colsample_bytree)
random.shuffle(min_child_weight)然后做融合。
ranking是先把每个模型的结果排序，然后再求平均。
比赛中发现，ranking比bagging（直接对每个模型结果求平均），略微好一丁点。
- uncleban (第一名)
大家久等的开源来了~~ https://github.com/uncleban/2019_cmb_fintech 第二题a榜第一baseline，这份代码可以复现线上a榜0.74151，b榜0.74242的成绩。提取特征部分用时2min左右，模型训练用时6min左右。 整理不易，大家的star就是最好的回报，谢谢大家，欢迎转发~
- VoyagerIII
B榜0.73611 交流分享，欢迎star 用的XGBoost 特征两百维，有一定注释。 https://github.com/VoyagerIII/CERES