工程能力UP | LightGBM的調參乾貨教程與並行優化

忽逢桃林發表於2020-07-23

並行優化

這是個人在競賽中對LGB模型進行調參的詳細過程記錄，主要包含下面六個步驟：

大學習率，確定估計器引數n_estimators/num_iterations/num_round/num_boost_round；
確定num_leaves和max_depth
確定min_data_in_leaf
確定bagging_fraction+bagging_freq和feature_fraction
確定L1L2正則reg_alpha和reg_lambda；
降低學習率

【這裡必須說一下，lightbg的引數的同義詞實在太多了，很多不同的參數列示的是同一個意思，不過本文中使用“/”分開】

0 並行優化

主要有兩種feature parallel特徵並行和data parallel資料並行。具體的過程我不也不瞭解，因為我沒有多個CPU給我耍（窮）。

feature parallel：每個worker有全部的訓練資料，但是他們只用部分特徵進行訓練，然後不同worker之間交流他們的區域性最優特徵和分裂點,比較出來哪一個是全域性最優的。
data parallel: 每一個worker有部分的樣本，然後繪製區域性特徵直方圖。彼此交流之後，得到全域性直方圖進行訓練。

【雖然具體的機制不太瞭解，但是最重要的是：小資料用feature parallel，大資料用data parallel】

1. 估計器數量

不管怎麼樣，我們先把學習率先定一個較高的值，這裡取 learning_rate = 0.1，其次確定估計器boosting/boost/boosting_type的型別，不過預設都會選gbdt。

這裡可以體現，雖然LGB和XGB經常拿來和GBDT比較，但是其本質都還是GBDT的boost思想

為了確定估計器的數目，也就是boosting迭代的次數，也可以說是殘差樹的數目，引數名為n_estimators/num_iterations/num_round/num_boost_round。我們可以先將該引數設成一個較大的數，然後在cv結果中檢視最優的迭代次數，具體如程式碼。

在這之前，我們必須給其他重要的引數一個初始值。初始值的意義不大，只是為了方便確定其他引數。下面先給定一下初始值：

以下引數根據具體專案要求定：

'boosting_type'/'boosting': 'gbdt'
'objective': 'regression'
'metric': 'rmse'

以下引數我選擇的初始值，你可以根據自己的情況來選擇：

'max_depth': 6     ###   根據問題來定咯，由於我的資料集不是很大，所以選擇了一個適中的值，其實4-10都無所謂。
'num_leaves': 50  ###   由於lightGBM是leaves_wise生長，官方說法是要小於2^max_depth
'subsample'/'bagging_fraction':0.8           ###  資料取樣
'colsample_bytree'/'feature_fraction': 0.8  ###  特徵取樣

下面我是用LightGBM的cv函式進行演示：

params = {
    'boosting_type': 'gbdt', 
    'objective': 'regression', 

    'learning_rate': 0.1, 
    'num_leaves': 50, 
    'max_depth': 6,

    'subsample': 0.8, 
    'colsample_bytree': 0.8, 
    }

data_train = lgb.Dataset(df_train, y_train, silent=True)
cv_results = lgb.cv(
    params, data_train, num_boost_round=1000, nfold=5, stratified=False, shuffle=True, metrics='rmse',
    early_stopping_rounds=50, verbose_eval=50, show_stdv=True, seed=0)

print('best n_estimators:', len(cv_results['rmse-mean']))
print('best cv score:', cv_results['rmse-mean'][-1])

執行結果是：

[50]	cv_agg's rmse: 1.38497 + 0.0202823
best n_estimators: 43
best cv score: 1.3838664241

所以我們得到了結果，在學習率0.1的時候，有43個估計器的時候效果最好。所以現在我們已經調整好了一個引數了：n_estimators/num_iterations/num_round/num_boost_round=43。

【在硬體裝置允許的條件下，學習率還是越小越好】

2. 提高擬合程度

這是提高精確度的最重要的引數。

max_depth:設定樹深度，深度越大可能過擬合
num_leaves:因為 LightGBM 使用的是 leaf-wise 的演算法，因此在調節樹的複雜程度時，使用的是 num_leaves 而不是 max_depth。大致換算關係：num_leaves = 2^(max_depth)，但是它的值的設定應該小於 2^(max_depth)，否則可能會導致過擬合。

【這裡雖然說了num_leaves與max_depth之間的關係，但是並不是嚴格的，大概在這個左右就好了。】

接下來同時對這兩個引數調優，引入sklearn中的GridSearchCV()函式進行網格搜尋，當然也可以使用貝葉斯搜尋，貝葉斯這個之前在個人部落格講過，之後我有空了再搬運到公眾號好了。

不過這個搜尋過程，非常耗時間，非常消耗精力。對於大資料集的話，建議貝葉斯，或者就簡單調整下就行了。一般這種引數優化的空間非常有限。

from sklearn.model_selection import GridSearchCV
### 我們可以建立lgb的sklearn模型，使用上面選擇的(學習率，評估器數目)
model_lgb = lgb.LGBMRegressor(objective='regression',num_leaves=50,
                              learning_rate=0.1, n_estimators=43, max_depth=6,
                              metric='rmse', bagging_fraction = 0.8,feature_fraction = 0.8)

params_test1={
    'max_depth': range(3,8,2),
    'num_leaves':range(50, 170, 30)
}
gsearch1 = GridSearchCV(estimator=model_lgb, param_grid=params_test1, scoring='neg_mean_squared_error', cv=5, verbose=1, n_jobs=4)

gsearch1.fit(df_train, y_train)
gsearch1.grid_scores_, gsearch1.best_params_, gsearch1.best_score_

來看下執行的結果：

Fitting 5 folds for each of 12 candidates, totalling 60 fits


[Parallel(n_jobs=4)]: Done  42 tasks      | elapsed:  2.0min
[Parallel(n_jobs=4)]: Done  60 out of  60 | elapsed:  3.1min finished


([mean: -1.88629, std: 0.13750, params: {'max_depth': 3, 'num_leaves': 50},
  mean: -1.88629, std: 0.13750, params: {'max_depth': 3, 'num_leaves': 80},
  mean: -1.88629, std: 0.13750, params: {'max_depth': 3, 'num_leaves': 110},
  mean: -1.88629, std: 0.13750, params: {'max_depth': 3, 'num_leaves': 140},
  mean: -1.86917, std: 0.12590, params: {'max_depth': 5, 'num_leaves': 50},
  mean: -1.86917, std: 0.12590, params: {'max_depth': 5, 'num_leaves': 80},
  mean: -1.86917, std: 0.12590, params: {'max_depth': 5, 'num_leaves': 110},
  mean: -1.86917, std: 0.12590, params: {'max_depth': 5, 'num_leaves': 140},
  mean: -1.89254, std: 0.10904, params: {'max_depth': 7, 'num_leaves': 50},
  mean: -1.86024, std: 0.11364, params: {'max_depth': 7, 'num_leaves': 80},
  mean: -1.86024, std: 0.11364, params: {'max_depth': 7, 'num_leaves': 110},
  mean: -1.86024, std: 0.11364, params: {'max_depth': 7, 'num_leaves': 140}],
 {'max_depth': 7, 'num_leaves': 80},
 -1.8602436718814157)

這裡執行了12個引數組合，得到的最優解是在max_depth為7，num_leaves為80的情況下，分數為-1.860。

這裡必須說一下，sklearn模型評估裡的scoring引數都是採用的higher return values are better than lower return values（較高的返回值優於較低的返回值）。

但是，我採用的metric策略採用的是均方誤差(rmse)，越低越好，所以sklearn就提供了neg_mean_squared_error引數，也就是返回metric的負數，所以就均方差來說，也就變成負數越大越好了。

所以，可以看到，最優解的分數為-1.860，轉化為均方差為np.sqrt(-(-1.860)) = 1.3639，明顯比step1的分數要好很多。(之前用的是rmse均方根誤差，要開方)

至此，我們將我們這步得到的最優解代入第三步。其實，我這裡只進行了粗調，如果要得到更好的效果，可以將max_depth在7附近多取幾個值，num_leaves在80附近多取幾個值。千萬不要怕麻煩，雖然這確實很麻煩。

params_test2={
    'max_depth': [6,7,8],
    'num_leaves':[68,74,80,86,92]
}

gsearch2 = GridSearchCV(estimator=model_lgb, param_grid=params_test2, scoring='neg_mean_squared_error', cv=5, verbose=1, n_jobs=4)
gsearch2.fit(df_train, y_train)
gsearch2.grid_scores_, gsearch2.best_params_, gsearch2.best_score_

Fitting 5 folds for each of 15 candidates, totalling 75 fits


[Parallel(n_jobs=4)]: Done  42 tasks      | elapsed:  2.8min
[Parallel(n_jobs=4)]: Done  75 out of  75 | elapsed:  5.1min finished


([mean: -1.87506, std: 0.11369, params: {'max_depth': 6, 'num_leaves': 68},
  mean: -1.87506, std: 0.11369, params: {'max_depth': 6, 'num_leaves': 74},
  mean: -1.87506, std: 0.11369, params: {'max_depth': 6, 'num_leaves': 80},
  mean: -1.87506, std: 0.11369, params: {'max_depth': 6, 'num_leaves': 86},
  mean: -1.87506, std: 0.11369, params: {'max_depth': 6, 'num_leaves': 92},
  mean: -1.86024, std: 0.11364, params: {'max_depth': 7, 'num_leaves': 68},
  mean: -1.86024, std: 0.11364, params: {'max_depth': 7, 'num_leaves': 74},
  mean: -1.86024, std: 0.11364, params: {'max_depth': 7, 'num_leaves': 80},
  mean: -1.86024, std: 0.11364, params: {'max_depth': 7, 'num_leaves': 86},
  mean: -1.86024, std: 0.11364, params: {'max_depth': 7, 'num_leaves': 92},
  mean: -1.88197, std: 0.11295, params: {'max_depth': 8, 'num_leaves': 68},
  mean: -1.89117, std: 0.12686, params: {'max_depth': 8, 'num_leaves': 74},
  mean: -1.86390, std: 0.12259, params: {'max_depth': 8, 'num_leaves': 80},
  mean: -1.86733, std: 0.12159, params: {'max_depth': 8, 'num_leaves': 86},
  mean: -1.86665, std: 0.12174, params: {'max_depth': 8, 'num_leaves': 92}],
 {'max_depth': 7, 'num_leaves': 68},
 -1.8602436718814157)

可見最大深度7是沒問題的，但是看細節的話，發現在最大深度為7的情況下，葉結點的數量對分數並沒有影響。

3. 降低過擬合

說到這裡，就該降低過擬合了。

min_data_in_leaf是一個很重要的引數, 也叫min_child_samples，它的值取決於訓練資料的樣本個樹和num_leaves. 將其設定的較大可以避免生成一個過深的樹, 但有可能導致欠擬合。
min_sum_hessian_in_leaf：也叫min_child_weight，使一個結點分裂的最小海森值之和，真拗口（Minimum sum of hessians in one leaf to allow a split. Higher values potentially decrease overfitting）

關於第二個引數，其實我不是非常的明白，因為不太瞭解hessian值和hessian矩陣？之後有空抽個時間好好學習一下，我學習的過程就是這樣查漏補缺2333。請大家關注公眾號，這樣不會錯過每一個乾貨

我們採用跟上面相同的方法進行：

params_test3={
    'min_child_samples': [18, 19, 20, 21, 22],
    'min_child_weight':[0.001, 0.002]
}
model_lgb = lgb.LGBMRegressor(objective='regression',num_leaves=80,
                              learning_rate=0.1, n_estimators=43, max_depth=7, 
                              metric='rmse', bagging_fraction = 0.8, feature_fraction = 0.8)
gsearch3 = GridSearchCV(estimator=model_lgb, param_grid=params_test3, scoring='neg_mean_squared_error', cv=5, verbose=1, n_jobs=4)
gsearch3.fit(df_train, y_train)
gsearch3.grid_scores_, gsearch3.best_params_, gsearch3.best_score_

結果是：

Fitting 5 folds for each of 10 candidates, totalling 50 fits


[Parallel(n_jobs=4)]: Done  42 tasks      | elapsed:  2.9min
[Parallel(n_jobs=4)]: Done  50 out of  50 | elapsed:  3.3min finished


([mean: -1.88057, std: 0.13948, params: {'min_child_samples': 18, 'min_child_weight': 0.001},
  mean: -1.88057, std: 0.13948, params: {'min_child_samples': 18, 'min_child_weight': 0.002},
  mean: -1.88365, std: 0.13650, params: {'min_child_samples': 19, 'min_child_weight': 0.001},
  mean: -1.88365, std: 0.13650, params: {'min_child_samples': 19, 'min_child_weight': 0.002},
  mean: -1.86024, std: 0.11364, params: {'min_child_samples': 20, 'min_child_weight': 0.001},
  mean: -1.86024, std: 0.11364, params: {'min_child_samples': 20, 'min_child_weight': 0.002},
  mean: -1.86980, std: 0.14251, params: {'min_child_samples': 21, 'min_child_weight': 0.001},
  mean: -1.86980, std: 0.14251, params: {'min_child_samples': 21, 'min_child_weight': 0.002},
  mean: -1.86750, std: 0.13898, params: {'min_child_samples': 22, 'min_child_weight': 0.001},
  mean: -1.86750, std: 0.13898, params: {'min_child_samples': 22, 'min_child_weight': 0.002}],
 {'min_child_samples': 20, 'min_child_weight': 0.001},
 -1.8602436718814157)

這是我經過粗調後細調的結果，可以看到，min_data_in_leaf的最優值為20，而min_sum_hessian_in_leaf對最後的值幾乎沒有影響。且這裡調參之後，最優的結果還是-1.86024，沒有提升。

4. 取樣降低過擬合

這兩個引數都是為了降低過擬合的。

feature_fraction引數來進行特徵的子抽樣。這個引數可以用來防止過擬合及提高訓練速度。

bagging_fraction+bagging_freq引數必須同時設定，bagging_fraction相當於subsample樣本取樣，可以使bagging更快的執行，同時也可以降擬合。bagging_freq預設0，表示bagging的頻率，0意味著沒有使用bagging，k意味著每k輪迭代進行一次bagging。

不同的引數，同樣的方法：

params_test4={
    'feature_fraction': [0.5, 0.6, 0.7, 0.8, 0.9],
    'bagging_fraction': [0.6, 0.7, 0.8, 0.9, 1.0]
}
model_lgb = lgb.LGBMRegressor(objective='regression',num_leaves=80,
                              learning_rate=0.1, n_estimators=43, max_depth=7, 
                              metric='rmse', bagging_freq = 5,  min_child_samples=20)
gsearch4 = GridSearchCV(estimator=model_lgb, param_grid=params_test4, scoring='neg_mean_squared_error', cv=5, verbose=1, n_jobs=4)
gsearch4.fit(df_train, y_train)
gsearch4.grid_scores_, gsearch4.best_params_, gsearch4.best_score_

Fitting 5 folds for each of 25 candidates, totalling 125 fits


[Parallel(n_jobs=4)]: Done  42 tasks      | elapsed:  2.6min
[Parallel(n_jobs=4)]: Done 125 out of 125 | elapsed:  7.1min finished


([mean: -1.90447, std: 0.15841, params: {'bagging_fraction': 0.6, 'feature_fraction': 0.5},
  mean: -1.90846, std: 0.13925, params: {'bagging_fraction': 0.6, 'feature_fraction': 0.6},
  mean: -1.91695, std: 0.14121, params: {'bagging_fraction': 0.6, 'feature_fraction': 0.7},
  mean: -1.90115, std: 0.12625, params: {'bagging_fraction': 0.6, 'feature_fraction': 0.8},
  mean: -1.92586, std: 0.15220, params: {'bagging_fraction': 0.6, 'feature_fraction': 0.9},
  mean: -1.88031, std: 0.17157, params: {'bagging_fraction': 0.7, 'feature_fraction': 0.5},
  mean: -1.89513, std: 0.13718, params: {'bagging_fraction': 0.7, 'feature_fraction': 0.6},
  mean: -1.88845, std: 0.13864, params: {'bagging_fraction': 0.7, 'feature_fraction': 0.7},
  mean: -1.89297, std: 0.12374, params: {'bagging_fraction': 0.7, 'feature_fraction': 0.8},
  mean: -1.89432, std: 0.14353, params: {'bagging_fraction': 0.7, 'feature_fraction': 0.9},
  mean: -1.88088, std: 0.14247, params: {'bagging_fraction': 0.8, 'feature_fraction': 0.5},
  mean: -1.90080, std: 0.13174, params: {'bagging_fraction': 0.8, 'feature_fraction': 0.6},
  mean: -1.88364, std: 0.14732, params: {'bagging_fraction': 0.8, 'feature_fraction': 0.7},
  mean: -1.88987, std: 0.13344, params: {'bagging_fraction': 0.8, 'feature_fraction': 0.8},
  mean: -1.87752, std: 0.14802, params: {'bagging_fraction': 0.8, 'feature_fraction': 0.9},
  mean: -1.88348, std: 0.13925, params: {'bagging_fraction': 0.9, 'feature_fraction': 0.5},
  mean: -1.87472, std: 0.13301, params: {'bagging_fraction': 0.9, 'feature_fraction': 0.6},
  mean: -1.88656, std: 0.12241, params: {'bagging_fraction': 0.9, 'feature_fraction': 0.7},
  mean: -1.89029, std: 0.10776, params: {'bagging_fraction': 0.9, 'feature_fraction': 0.8},
  mean: -1.88719, std: 0.11915, params: {'bagging_fraction': 0.9, 'feature_fraction': 0.9},
  mean: -1.86170, std: 0.12544, params: {'bagging_fraction': 1.0, 'feature_fraction': 0.5},
  mean: -1.87334, std: 0.13099, params: {'bagging_fraction': 1.0, 'feature_fraction': 0.6},
  mean: -1.85412, std: 0.12698, params: {'bagging_fraction': 1.0, 'feature_fraction': 0.7},
  mean: -1.86024, std: 0.11364, params: {'bagging_fraction': 1.0, 'feature_fraction': 0.8},
  mean: -1.87266, std: 0.12271, params: {'bagging_fraction': 1.0, 'feature_fraction': 0.9}],
 {'bagging_fraction': 1.0, 'feature_fraction': 0.7},
 -1.8541224387666373)

從這裡可以看出來，bagging_feaction和feature_fraction的理想值分別是1.0和0.7，一個很重要原因就是，我的樣本數量比較小(4000+)，但是特徵數量很多(1000+)。所以，這裡我們取更小的步長，對feature_fraction進行更細緻的取值。

下面微調一下：

params_test5={
    'feature_fraction': [0.62, 0.65, 0.68, 0.7, 0.72, 0.75, 0.78 ]
}
model_lgb = lgb.LGBMRegressor(objective='regression',num_leaves=80,
                              learning_rate=0.1, n_estimators=43, max_depth=7, 
                              metric='rmse',  min_child_samples=20)
gsearch5 = GridSearchCV(estimator=model_lgb, param_grid=params_test5, scoring='neg_mean_squared_error', cv=5, verbose=1, n_jobs=4)
gsearch5.fit(df_train, y_train)
gsearch5.grid_scores_, gsearch5.best_params_, gsearch5.best_score_

Fitting 5 folds for each of 7 candidates, totalling 35 fits


[Parallel(n_jobs=4)]: Done  35 out of  35 | elapsed:  2.3min finished


([mean: -1.86696, std: 0.12658, params: {'feature_fraction': 0.62},
  mean: -1.88337, std: 0.13215, params: {'feature_fraction': 0.65},
  mean: -1.87282, std: 0.13193, params: {'feature_fraction': 0.68},
  mean: -1.85412, std: 0.12698, params: {'feature_fraction': 0.7},
  mean: -1.88235, std: 0.12682, params: {'feature_fraction': 0.72},
  mean: -1.86329, std: 0.12757, params: {'feature_fraction': 0.75},
  mean: -1.87943, std: 0.12107, params: {'feature_fraction': 0.78}],
 {'feature_fraction': 0.7},
 -1.8541224387666373)

好吧，feature_fraction就是0.7了

5. L1與L2

正則化引數lambda_l1(reg_alpha), lambda_l2(reg_lambda)，毫無疑問，是降低過擬合的，兩者分別對應l1正則化和l2正則化。我們也來嘗試一下使用這兩個引數。

params_test6={
    'reg_alpha': [0, 0.001, 0.01, 0.03, 0.08, 0.3, 0.5],
    'reg_lambda': [0, 0.001, 0.01, 0.03, 0.08, 0.3, 0.5]
}
model_lgb = lgb.LGBMRegressor(objective='regression',num_leaves=80,
                              learning_rate=0.b1, n_estimators=43, max_depth=7, 
                              metric='rmse',  min_child_samples=20, feature_fraction=0.7)
gsearch6 = GridSearchCV(estimator=model_lgb, param_grid=params_test6, scoring='neg_mean_squared_error', cv=5, verbose=1, n_jobs=4)
gsearch6.fit(df_train, y_train)
gsearch6.grid_scores_, gsearch6.best_params_, gsearch6.best_score_

Fitting 5 folds for each of 49 candidates, totalling 245 fits


[Parallel(n_jobs=4)]: Done  42 tasks      | elapsed:  2.8min
[Parallel(n_jobs=4)]: Done 192 tasks      | elapsed: 10.6min
[Parallel(n_jobs=4)]: Done 245 out of 245 | elapsed: 13.3min finished


([mean: -1.85412, std: 0.12698, params: {'reg_alpha': 0, 'reg_lambda': 0},
  mean: -1.85990, std: 0.13296, params: {'reg_alpha': 0, 'reg_lambda': 0.001},
  mean: -1.86367, std: 0.13634, params: {'reg_alpha': 0, 'reg_lambda': 0.01},
  mean: -1.86787, std: 0.13881, params: {'reg_alpha': 0, 'reg_lambda': 0.03},
  mean: -1.87099, std: 0.12476, params: {'reg_alpha': 0, 'reg_lambda': 0.08},
  mean: -1.87670, std: 0.11849, params: {'reg_alpha': 0, 'reg_lambda': 0.3},
  mean: -1.88278, std: 0.13064, params: {'reg_alpha': 0, 'reg_lambda': 0.5},
  mean: -1.86190, std: 0.13613, params: {'reg_alpha': 0.001, 'reg_lambda': 0},
  mean: -1.86190, std: 0.13613, params: {'reg_alpha': 0.001, 'reg_lambda': 0.001},
  mean: -1.86515, std: 0.14116, params: {'reg_alpha': 0.001, 'reg_lambda': 0.01},
  mean: -1.86908, std: 0.13668, params: {'reg_alpha': 0.001, 'reg_lambda': 0.03},
  mean: -1.86852, std: 0.12289, params: {'reg_alpha': 0.001, 'reg_lambda': 0.08},
  mean: -1.88076, std: 0.11710, params: {'reg_alpha': 0.001, 'reg_lambda': 0.3},
  mean: -1.88278, std: 0.13064, params: {'reg_alpha': 0.001, 'reg_lambda': 0.5},
  mean: -1.87480, std: 0.13889, params: {'reg_alpha': 0.01, 'reg_lambda': 0},
  mean: -1.87284, std: 0.14138, params: {'reg_alpha': 0.01, 'reg_lambda': 0.001},
  mean: -1.86030, std: 0.13332, params: {'reg_alpha': 0.01, 'reg_lambda': 0.01},
  mean: -1.86695, std: 0.12587, params: {'reg_alpha': 0.01, 'reg_lambda': 0.03},
  mean: -1.87415, std: 0.13100, params: {'reg_alpha': 0.01, 'reg_lambda': 0.08},
  mean: -1.88543, std: 0.13195, params: {'reg_alpha': 0.01, 'reg_lambda': 0.3},
  mean: -1.88076, std: 0.13502, params: {'reg_alpha': 0.01, 'reg_lambda': 0.5},
  mean: -1.87729, std: 0.12533, params: {'reg_alpha': 0.03, 'reg_lambda': 0},
  mean: -1.87435, std: 0.12034, params: {'reg_alpha': 0.03, 'reg_lambda': 0.001},
  mean: -1.87513, std: 0.12579, params: {'reg_alpha': 0.03, 'reg_lambda': 0.01},
  mean: -1.88116, std: 0.12218, params: {'reg_alpha': 0.03, 'reg_lambda': 0.03},
  mean: -1.88052, std: 0.13585, params: {'reg_alpha': 0.03, 'reg_lambda': 0.08},
  mean: -1.87565, std: 0.12200, params: {'reg_alpha': 0.03, 'reg_lambda': 0.3},
  mean: -1.87935, std: 0.13817, params: {'reg_alpha': 0.03, 'reg_lambda': 0.5},
  mean: -1.87774, std: 0.12477, params: {'reg_alpha': 0.08, 'reg_lambda': 0},
  mean: -1.87774, std: 0.12477, params: {'reg_alpha': 0.08, 'reg_lambda': 0.001},
  mean: -1.87911, std: 0.12027, params: {'reg_alpha': 0.08, 'reg_lambda': 0.01},
  mean: -1.86978, std: 0.12478, params: {'reg_alpha': 0.08, 'reg_lambda': 0.03},
  mean: -1.87217, std: 0.12159, params: {'reg_alpha': 0.08, 'reg_lambda': 0.08},
  mean: -1.87573, std: 0.14137, params: {'reg_alpha': 0.08, 'reg_lambda': 0.3},
  mean: -1.85969, std: 0.13109, params: {'reg_alpha': 0.08, 'reg_lambda': 0.5},
  mean: -1.87632, std: 0.12398, params: {'reg_alpha': 0.3, 'reg_lambda': 0},
  mean: -1.86995, std: 0.12651, params: {'reg_alpha': 0.3, 'reg_lambda': 0.001},
  mean: -1.86380, std: 0.12793, params: {'reg_alpha': 0.3, 'reg_lambda': 0.01},
  mean: -1.87577, std: 0.13002, params: {'reg_alpha': 0.3, 'reg_lambda': 0.03},
  mean: -1.87402, std: 0.13496, params: {'reg_alpha': 0.3, 'reg_lambda': 0.08},
  mean: -1.87032, std: 0.12504, params: {'reg_alpha': 0.3, 'reg_lambda': 0.3},
  mean: -1.88329, std: 0.13237, params: {'reg_alpha': 0.3, 'reg_lambda': 0.5},
  mean: -1.87196, std: 0.13099, params: {'reg_alpha': 0.5, 'reg_lambda': 0},
  mean: -1.87196, std: 0.13099, params: {'reg_alpha': 0.5, 'reg_lambda': 0.001},
  mean: -1.88222, std: 0.14735, params: {'reg_alpha': 0.5, 'reg_lambda': 0.01},
  mean: -1.86618, std: 0.14006, params: {'reg_alpha': 0.5, 'reg_lambda': 0.03},
  mean: -1.88579, std: 0.12398, params: {'reg_alpha': 0.5, 'reg_lambda': 0.08},
  mean: -1.88297, std: 0.12307, params: {'reg_alpha': 0.5, 'reg_lambda': 0.3},
  mean: -1.88148, std: 0.12622, params: {'reg_alpha': 0.5, 'reg_lambda': 0.5}],
 {'reg_alpha': 0, 'reg_lambda': 0},
 -1.8541224387666373)

哈哈，看來我多此一舉了。

6. 降低learning_rate

回到第一步，不過我們使用的是已經優化好的引數值：

params = {
    'boosting_type': 'gbdt', 
    'objective': 'regression', 

    'learning_rate': 0.005, 
    'num_leaves': 80, 
    'max_depth': 7,
    'min_data_in_leaf': 20,

    'subsample': 1, 
    'colsample_bytree': 0.7, 
    }

data_train = lgb.Dataset(df_train, y_train, silent=True)
cv_results = lgb.cv(
    params, data_train, num_boost_round=10000, nfold=5, stratified=False, shuffle=True, metrics='rmse',
    early_stopping_rounds=50, verbose_eval=100, show_stdv=True)

print('best n_estimators:', len(cv_results['rmse-mean']))
print('best cv score:', cv_results['rmse-mean'][-1])

[100]	cv_agg's rmse: 1.52939 + 0.0261756
[200]	cv_agg's rmse: 1.43535 + 0.0187243
[300]	cv_agg's rmse: 1.39584 + 0.0157521
[400]	cv_agg's rmse: 1.37935 + 0.0157429
[500]	cv_agg's rmse: 1.37313 + 0.0164503
[600]	cv_agg's rmse: 1.37081 + 0.0172752
[700]	cv_agg's rmse: 1.36942 + 0.0177888
[800]	cv_agg's rmse: 1.36854 + 0.0180575
[900]	cv_agg's rmse: 1.36817 + 0.0188776
[1000]	cv_agg's rmse: 1.36796 + 0.0190279
[1100]	cv_agg's rmse: 1.36783 + 0.0195969
best n_estimators: 1079
best cv score: 1.36772351783

參考連結：

貝葉斯全域性優化（LightGBM調參）
2019-08-20
優化
LightGBM核心解析與調參
2019-03-03
乾貨收藏 | 如何優化前端效能？
2020-10-16
優化前端
乾貨！MySQL大表優化方案（1）
2018-10-17
MySql優化
乾貨｜優酷小程式優化實戰
2021-11-12
優化
乾貨｜智慧編譯優化框架TACT
2017-12-27
編譯優化框架
程式分析與優化 - 10 指令級並行
2022-07-09
優化並行
【真·乾貨】MySQL 索引及優化實戰
2018-10-31
MySql索引優化
用Dask並行化特徵工程！
2018-08-20
並行特徵工程
前端工程與效能優化
2015-09-05
前端優化
乾貨！谷歌推薦的技術能力提升指南
2015-05-15
谷歌
【乾貨分享】20條PHP程式碼優化技巧
2015-05-19
PHP優化
SQL效能優化的祕訣，快來圍觀，乾貨！
2021-01-05
SQL優化
Nginx安全優化與效能調優
2020-07-25
Nginx優化
讓Elasticsearch飛起來!——效能優化實踐乾貨
2019-02-21
Elasticsearch優化
讓 Elasticsearch 飛起來!——效能優化實踐乾貨
2019-01-01
Elasticsearch優化
網易影片雲：MySQL 5.7 並行複製實現原理與調優
2017-07-20
MySql並行
乾貨：成為優秀的JAVA工程師必讀的經典書籍
2018-12-16
Java工程師
Buffer cache 的調整與優化(二)
2011-07-29
優化
Buffer cache 的調整與優化(一)
2011-07-29
優化
大分割槽表的手工並行優化
2013-11-08
並行優化
【乾貨】快速安裝 GitLab 並漢化
2017-12-07
Gitlab
乾貨分享：螞蟻金服前端框架和工程化實踐
2019-07-17
前端框架
【Mysql】mysql5.7新特性之-並行複製實現原理與調優
2015-11-12
MySql並行
乾貨 |如何優雅的在手機上進行Python程式設計
2018-11-02
Python程式設計
Parallel Query Tuning(並行化查詢優化)
2010-05-06
Parallel並行優化
SQL調優日記：並行等待的原理和問題排查
2016-10-28
SQL並行
TiDB 查詢優化及調優系列（四）查詢執行計劃的調整及優化原理
2022-05-24
TiDB優化
Oracle調優---調整優化器行為的一般性原則
2007-11-26
Oracle優化
SQL優化參考
2018-09-15
SQL優化
sql 優化參考
2013-02-03
SQL優化
乾貨好文:分散式資料庫DDL的編譯與執行
2022-06-17
分散式資料庫編譯
網雲穿內網穿透乾貨教程
2019-12-24
內網穿透
Java程式效能調優阿姆達爾定律、快取元件、並行開發、執行緒池、JVM調優
2014-08-05
Java快取元件並行執行緒JVM
挑戰月薪30K | 前端效能優化的12 條建議（乾貨收藏）
2020-12-22
前端優化
並行工程的本質分析（轉）
2007-08-15
並行
乾貨丨RPA工程中的資料處理問題
2019-08-22
探索：優雅地實現非同步方法的並行化
2023-02-09
非同步並行