ML之lightgbm.sklearn：LGBMClassifier函数的简介、具体案例、调参技巧之详细攻略

今天药忘吃喽~ 2022-12-17 14:51 886阅读 0赞

**目录**

LGBMClassifier函数的简介、具体案例、调参技巧

LGBMClassifier函数的调参技巧

1、lightGBM适合较大数据集的样本

2、建议使用更小的learning\_rate和更大的num\_iteration

3、样本不平衡调参技巧

4、调参时，可将参数字典分为两大类

LGBMClassifier函数简介

1、所有弱学习器的参数

2、具体函数解释

--------------------

# LGBMClassifier函数的简介、具体案例、调参技巧 #

## LGBMClassifier函数的调参技巧 ##

### 1、lightGBM适合较大数据集的样本 ###

而对于较小的数据集(<10000条记录)，lightGBM可能不是最佳选择。所以，如果进行调优lightgbm参数，这可能没有帮助。

### 2、建议使用更小的learning\_rate和更大的num\_iteration ###

此外，如果您想要更高的num\_iteration，那么您应该使用early\_stopping\_rounds，以便在无法学习任何有用的内容时停止训练。

### 3、样本不平衡调参技巧 ###

lightgbm中，可以设置两个参数is\_unbalance和scale\_pos\_weight。  
is\_unbalace：当其为True时，算法将尝试自动平衡占主导地位的标签的权重(使用列集中的pos/neg分数)  
scale\_pos\_weight：默认1，即假设正负标签都是相等的。在不平衡数据集的情况下，建议使用以下公式：  
sample\_pos\_weight = number of negative samples / number of positive samples

### 4、调参时，可将参数字典分为两大类 ###

[https://sites.google.com/view/lauraepp/parameters][https_sites.google.com_view_lauraepp_parameters]

<table> 
 <tbody> 
 <tr> 
 <td style="vertical-align:top;width:73.25pt;"> 调优参数 </td> 
 <td style="vertical-align:top;width:369.55pt;"> search_params = {'learning_rate': 0.4, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;'max_depth': 15, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;'num_leaves': 20, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;'feature_fraction': 0.8, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;'subsample': 0.2} </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:73.25pt;"> 固定参数 </td> 
 <td style="vertical-align:top;width:369.55pt;"> fixed_params={'objective': 'binary', &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;'metric': 'auc', &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;'is_unbalance':True, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;'boosting':'gbdt', &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;'num_boost_round':300, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;'early_stopping_rounds':30} </td> 
 </tr> 
 </tbody> 
</table>

## LGBMClassifier函数简介 ##

**LightGBM原论文**：[https://papers.nips.cc/paper/6907-lightgbm-a-highly-efficient-gradient-boosting-decision-tree][https_papers.nips.cc_paper_6907-lightgbm-a-highly-efficient-gradient-boosting-decision-tree]

### 1、所有弱学习器的参数 ###

<table> 
 <tbody> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> 参数及其默认 </td> 
 <td style="vertical-align:top;"> 参数解释及调参技巧 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> boosting_type='gbdt' </td> 
 <td style="vertical-align:top;"> 提升树的类型，常用的梯度提升方法包括gbdt、dart、goss、rf。可以尝试运行不同类型的渐变增强提升方法。 (1)、gbdt：这是传统的梯度提升决策树，也是基于XGBoost和pGBRT等优秀库背后的算法。gbdt精度高、效率高、稳定性好，目前已得到广泛的应用。但是，它的主要缺点是，在每个树节点中找到最佳分割点非常耗时，而且会消耗内存。下边其它的提升方法试图解决这个问题。 (2)、dart：即Dropouts meet Multiple Additive Regression Trees，dart利用dropout技巧(源自深度神经网络)解决过拟合的Regression Trees，改进模型正则化。gbdt存在过度专门化（over-specialization）的问题，这意味着在以后的迭代中添加的树往往只会影响对少数实例的预测，而对其余实例的贡献则可以忽略不计。添加dropout会使树在以后的迭代中更加难以专门化那些少数的示例，从而提高性能。 它的原理是随机丢弃生成的决策树，然后再从剩下的决策树集中迭代优化提升树。它的特点是 
 <ul> 
 <li>训练慢：因为随机dropout不使用用于保存预测结果的buffer，所以训练会更慢。</li> 
 <li>随机导致不稳定：因为随机，早停可能不够稳定。</li> 
 </ul>dart与gbdt的不同点：计算下一棵树要拟合的梯度的时，仅仅随机从已经生成的树中选取一部分。 注意dart添加一棵树时需要先归一化。 (3)、goss ：基于梯度的单边采样，该方法命名为lightgbm的最重要原因就是其使用了基于Goss方法。goss的基本思想是首先对训练集数据根据梯度排序，预设一个比例划分梯度大小，保留在所有样本中梯度大的数据样本；再设置一个采样比例，从梯度小的样本中按比例抽取样本。为了弥补对样本分布造成的影响，goss算法在计算信息增益时，会对较小梯度的数据集乘以一个系数，用来放大。这样，在计算信息增益时，算法可以更加关注“未被充分训练”的样本数据。 &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;goss通过对较小的样本数据集估算增益，大大的减少了计算量。而且通过证明，goss算法不会过多的降低训练的精度。 标准的gbdt是可靠的，但在大型数据集上速度不够快。因此goss提出了一种基于梯度的采样方法来避免搜索整个搜索空间。其实，对于每个数据实例，当梯度很小时，这意味着不用担心数据是经过良好训练的，而当梯度很大时，应该重新训练。数据实例有大的和小的渐变。因此，goss以一个大的梯度保存所有数据，并对一个小梯度的数据进行随机抽样(这就是为什么它被称为单边抽样)。这使得搜索空间更小，goss的收敛速度更快。 (4)、rf&nbsp;：随机森林。切记，如果将增强设置为rf，那么lightgbm算法表现为随机森林而不是增强树。根据文档可知，要使用rf，必须使用bagging_fraction和feature_fraction小于1。 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> objective ='binary' </td> 
 <td style="vertical-align:top;"> 目标，"regression"、"binary" </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> metric='binary_logloss' </td> 
 <td style="vertical-align:top;"> 模型度量标准，"rmse"、"auc"、'binary_logloss' </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> learning_rate=0.1 </td> 
 <td style="vertical-align:top;"> 学习率 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> n_estimators=10 </td> 
 <td style="vertical-align:top;"> 拟合的树的棵树，可以理解为训练的轮数。 弱学习器的个数，其中gbdt原理是利用通过梯度不断拟合新的弱学习器，直到达到设定的弱学习器的数量。 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> max_depth=-1 </td> 
 <td style="vertical-align:top;"> 最大树的深度。每个弱学习器也就是决策树的最大深度。 其中，-1表示不限制。 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> num_leavel=32 </td> 
 <td style="vertical-align:top;"> 树的最大叶子数，控制模型复杂性的最重要参数之一。对比在xgboost中，一般为2^(max_depth) 因LightGBM使用的是leaf-wise的算法，因此在调节树的复杂程度时，使用的是num_leaves而不是max_depth。它们之间大致换算关系：num_leaves = 2^(max_depth)。即它的值的设置应该小于2^(max_depth)，否则会进行警告，可能会导致过拟合。 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> bagging_freq=15 </td> 
 <td style="vertical-align:top;"> 控制过拟合。 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> bagging_fraction= 0.8 </td> 
 <td style="vertical-align:top;"> 子样例，来控制过拟合。 可以指定每个树构建迭代使用的行数百分比。这意味着将随机选择一些行来匹配每个学习者(树)。这不仅提高了泛化能力，也提高了训练速度。 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> feature_fraction=0.8 </td> 
 <td style="vertical-align:top;"> 子特征处理列采样，来控制过拟合。 它将在每次迭代(树)上随机选择特征子集。例如，如果将其设置为0.8，它将在训练每棵树之前选择60%的特性。它常用来加速训练和处理过拟合。 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> subsample=1.0 </td> 
 <td style="vertical-align:top;"> 训练样本采样率，行 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> colsample_bytree=1.0&nbsp; </td> 
 <td style="vertical-align:top;"> 训练特征采样率，列 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> subsample_freq=1 </td> 
 <td style="vertical-align:top;"> 子样本频率 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> reg_alpha=0.5 </td> 
 <td style="vertical-align:top;"> L1正则化系数 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> reg_lambda=0.5 </td> 
 <td style="vertical-align:top;"> L2正则化系数 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> min_split_gain=0.0 </td> 
 <td style="vertical-align:top;"> 最小分割增益 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> min_child_weight=0.001&nbsp; </td> 
 <td style="vertical-align:top;"> 分支结点的最小权重 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> min_child_samples=20 </td> 
 <td style="vertical-align:top;"> &nbsp; </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> random_state=None </td> 
 <td style="vertical-align:top;"> 随机种子数 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> n_jobs=-1 </td> 
 <td style="vertical-align:top;"> 并行运行多线程核心数 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> silent=True </td> 
 <td style="vertical-align:top;"> 训练过程是否打印日志信息 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:178px;"> verbose=-1 </td> 
 <td style="vertical-align:top;"> &nbsp; </td> 
 </tr> 
 </tbody> 
</table>

### 2、具体函数解释 ###

<table> 
 <tbody> 
 <tr> 
 <td style="vertical-align:top;width:471px;"> class LGBMClassifier&nbsp;Found at: lightgbm.sklearn &nbsp; class LGBMClassifier(LGBMModel, _LGBMClassifierBase): &nbsp;&nbsp;&nbsp;&nbsp;"""LightGBM classifier.""" &nbsp;&nbsp;&nbsp;&nbsp;def fit(self, X, y, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;sample_weight=None, init_score=None, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;eval_set=None, eval_names=None, eval_sample_weight=None, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;eval_class_weight=None, eval_init_score=None, eval_metric=None, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;early_stopping_rounds=None, verbose=True, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;feature_name='auto', categorical_feature='auto', callbacks=None): &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;"""Docstring is inherited from the LGBMModel.""" &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;_LGBMAssertAllFinite(y) &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;_LGBMCheckClassificationTargets(y) &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;self._le = _LGBMLabelEncoder().fit(y) &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;_y = self._le.transform(y) &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;self._classes = self._le.classes_ &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;self._n_classes = len(self._classes) &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;if self._n_classes &gt; 2: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;# Switch to using a multiclass objective in the underlying LGBM &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;instance &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;ova_aliases = "multiclassova", "multiclass_ova", "ova", "ovr" &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;if self._objective not in ova_aliases and not callable(self. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;_objective): &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;self._objective = "multiclass" &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;if eval_metric in ('logloss', 'binary_logloss'): &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;eval_metric = "multi_logloss" &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;elif eval_metric in ('error', 'binary_error'): &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;eval_metric = "multi_error" &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;elif eval_metric in ('logloss', 'multi_logloss'): &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;eval_metric = 'binary_logloss' &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;elif eval_metric in ('error', 'multi_error'): &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;eval_metric = 'binary_error' &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;if eval_set is not None: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;if isinstance(eval_set, tuple): &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;eval_set = [eval_set] &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;for i, (valid_x, valid_y) in enumerate(eval_set): &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;if valid_x is X and valid_y is y: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;eval_set[i] = valid_x, _y &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;else: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;eval_set[i] = valid_x, self._le.transform(valid_y) &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;super(LGBMClassifier, self).fit(X, _y, sample_weight=sample_weight, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;init_score=init_score, eval_set=eval_set, eval_names=eval_names, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;eval_sample_weight=eval_sample_weight, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;eval_class_weight=eval_class_weight, eval_init_score=eval_init_score, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;eval_metric=eval_metric, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;early_stopping_rounds=early_stopping_rounds, verbose=verbose, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;feature_name=feature_name, categorical_feature=categorical_feature, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;callbacks=callbacks) &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;return self &nbsp;&nbsp;&nbsp;&nbsp; &nbsp;&nbsp;&nbsp;&nbsp;fit.__doc__ = LGBMModel.fit.__doc__ &nbsp;&nbsp;&nbsp;&nbsp;def predict(self, X, raw_score=False, num_iteration=None, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;pred_leaf=False, pred_contrib=False, **kwargs): &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;"""Docstring is inherited from the LGBMModel.""" &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;result = self.predict_proba(X, raw_score, num_iteration, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;pred_leaf, pred_contrib, **kwargs) &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;if raw_score or pred_leaf or pred_contrib: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;return result &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;else: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;class_index = np.argmax(result, axis=1) &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;return self._le.inverse_transform(class_index) &nbsp;&nbsp;&nbsp;&nbsp; &nbsp;&nbsp;&nbsp;&nbsp;predict.__doc__ = LGBMModel.predict.__doc__ &nbsp;&nbsp;&nbsp;&nbsp;def predict_proba(self, X, raw_score=False, num_iteration=None, &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;pred_leaf=False, pred_contrib=False, **kwargs): &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;"""Return the predicted probability for each class for each sample.&nbsp;返回每个类和每个样本的预测概率。 </td> 
 <td style="vertical-align:top;width:433px;"> &nbsp; </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:471px;"> &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Parameters &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;---------- &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;X&nbsp;: array-like or sparse matrix of shape = [n_samples, n_features]&nbsp;Input features matrix. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;raw_score&nbsp;: bool, optional&nbsp;(default=False). Whether to predict raw scores. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;num_iteration&nbsp;: int or None, optional (default=None). Limit number of iterations in the prediction.&nbsp;If None, if the best iteration exists, it is used; otherwise, all trees are used.&nbsp;&nbsp;If &lt;= 0, all trees are used (no limits). &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;pred_leaf&nbsp;: bool, optional (default=False). Whether to predict leaf index. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;pred_contrib&nbsp;: bool, optional (default=False). Whether to predict feature contributions. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Note &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;---- &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;If you want to get more explanation for your model's predictions using SHAP values&nbsp;like SHAP interaction values,&nbsp;you can install shap package (https://github.com/slundberg/shap). &nbsp; &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;**kwargs &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Other parameters for the prediction. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Returns &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;------- &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;predicted_probability&nbsp;: array-like of shape = [n_samples, n_classes]. The predicted probability for each class for each sample. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;X_leaves&nbsp;: array-like of shape = [n_samples, n_trees * n_classes]. If ``pred_leaf=True``, the predicted leaf every tree for each sample. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;X_SHAP_values&nbsp;: array-like of shape = [n_samples, (n_features + 1) *n_classes]. If ``pred_contrib=True``, the each feature contributions for each sample. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;""" &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;result = super(LGBMClassifier, self).predict(X, raw_score, num_iteration,&nbsp;pred_leaf, pred_contrib, **kwargs). if self._n_classes &gt; 2 or pred_leaf or pred_contrib: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;return result &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;else: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;return np.vstack((1. - result, result)).transpose() </td> 
 <td style="vertical-align:top;width:433px;"> 参数 ---------- X:&nbsp;类数组形状或稀疏矩阵= [n_samples, n_features]输入特征矩阵。 raw_score&nbsp;: bool，可选(默认=False)。是否预测原始分数。 num_iteration:&nbsp;&nbsp;int或None，可选(默认=None)。预测中的极限迭代次数。如果没有，如果存在最好的迭代，就使用它;否则，使用所有树。如果&lt;= 0，则使用所有树(没有限制)。 pred_leaf: &nbsp;bool，可选(默认=False)。是否预测叶指数。 pred_contrib: &nbsp;bool，可选(默认=False)。是否预测特性贡献。 请注意 ---- 如果您想使用SHAP值(如SHAP交互值)对模型的预测进行更多的解释，您可以安装SHAP包(https://github.com/slundberg/shap)。 &nbsp; * * kwargs 用于预测的其他参数。 返回 ------- predicted_probability&nbsp;: 类数组形状= [n_samples, n_classes]。每个类，每个样本的预测概率。 X_leaves:&nbsp;类数组的形状= [n_samples, n_trees * n_classes]。如果' ' pred_leaf=True ' '，则为每个样本的每棵树预测的叶。 X_SHAP_values:&nbsp;类数组形状= [n_samples， (n_features + 1) *n_classes]。如果' ' pred_contrib=True ' '，则每个特性为每个示例贡献。 &nbsp; """ &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;result = super(LGBMClassifier, self).predict(X, raw_score, num_iteration,&nbsp;pred_leaf, pred_contrib, **kwargs). if self._n_classes &gt; 2 or pred_leaf or pred_contrib: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;return result &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;else: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;return np.vstack((1. - result, result)).transpose() </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;width:471px;"> &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;@property &nbsp;&nbsp;&nbsp;&nbsp;def classes_(self): &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;"""Get the class label array.""" &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;if self._classes is None: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;raise LGBMNotFittedError('No classes found. Need to call fit &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;beforehand.') &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;return self._classes &nbsp;&nbsp;&nbsp;&nbsp; &nbsp;&nbsp;&nbsp;&nbsp;@property &nbsp;&nbsp;&nbsp;&nbsp;def n_classes_(self): &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;"""Get the number of classes.""" &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;if self._n_classes is None: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;raise LGBMNotFittedError('No classes found. Need to call fit &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;beforehand.') &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;return self._n_classes </td> 
 <td style="vertical-align:top;width:433px;"> &nbsp; </td> 
 </tr> 
 </tbody> 
</table>

[https_sites.google.com_view_lauraepp_parameters]: https://sites.google.com/view/lauraepp/parameters
[https_papers.nips.cc_paper_6907-lightgbm-a-highly-efficient-gradient-boosting-decision-tree]: https://papers.nips.cc/paper/6907-lightgbm-a-highly-efficient-gradient-boosting-decision-tree