梯度提升树回归
概述
“梯度提升树回归”节点用于生成回归模型，是一种基于决策树的迭代回归算法。该算法采用迭代的思想不断地构建决策树模型，每棵树都是通过梯度优化损失函数而构建，从而达到从基准值到目标值的逼近。算法思想可简单理解成：后一次模型都是针对前一次模型预测出错的情况进行修正，模型随着迭代不断地改进，从而获得比较好的预测效果。
梯度提升树回归的损失函数为均方差损失函数，如下所示：
其中，N 表示样本数量，xi 表示样本i 的特征，yi 表示样本i 的标签，F(xi) 表示样本i 预测的标签。
输入
参数 |
子参数 |
参数说明 |
|---|---|---|
inputs |
dataframe |
inputs为字典类型，dataframe为pyspark中的DataFrame类型对象 |
输出
spark pipeline类型的模型
参数说明
参数 |
子参数 |
参数说明 |
|---|---|---|
b_use_default_encoder |
- |
是否使用默认编码，默认为True |
input_features_str |
- |
输入的列名以逗号分隔组成的字符串，例如： "column_a" "column_a,column_b" |
label_col |
- |
目标列 |
regressor_feature_vector_col |
- |
算子输入的特征向量列的列名，默认为"model_features" |
max_depth |
- |
树的最大深度，默认为5 |
max_bins |
- |
最大分箱数，默认为32 |
min_instances_per_node |
- |
节点分割时，要求子节点必须包含的最少实例数，默认为1 |
min_info_gain |
- |
节点是否分割要求的最小信息增益，默认为0.0 |
subsampling_rate |
- |
学习每棵决策树用到的训练集的抽样比例，默认为1.0 |
loss_type |
- |
损失函数类型，支持squared、absolute，默认为"squared" |
max_iter |
- |
最大迭代次数，默认为20 |
step_size |
- |
步长，默认为0.1 |
样例
inputs = {
"dataframe": None # @input {"label":"dataframe","type":"DataFrame"}
}
params = {
"inputs": inputs,
"b_output_action": True,
"b_use_default_encoder": True, # @param {"label": "b_use_default_encoder", "type": "boolean", "required": "true", "helpTip": ""}
"input_features_str": "", # @param {"label": "input_features_str", "type": "string", "required": "false", "helpTip": ""}
"outer_pipeline_stages": None,
"label_col": "", # @param {"label": "label_col", "type": "string", "required": "true", "helpTip": "target label column"}
"regressor_feature_vector_col": "model_features", # @param {"label": "regressor_feature_vector_col", "type": "string", "required": "true", "helpTip": ""}
"max_depth": 5, # @param {"label": "max_depth", "type": "integer", "required": "true", "range": "(0,2147483647]", "helpTip": ""}
"max_bins": 32, # @param {"label": "max_bins", "type": "integer", "required": "true", "range": "(0,2147483647]", "helpTip": ""}
"min_instances_per_node": 1, # @param {"label": "min_instances_per_node", "type": "integer", "required": "true", "range": "(0,2147483647]", "helpTip": ""}
"min_info_gain": 0.0, # @param {"label": "min_info_gain", "type": "number", "required": "true", "range": "[0.0,none)", "helpTip": ""}
"subsampling_rate": 1.0, # @param {"label": "subsampling_rate", "type": "number", "required": "true", "range": "(0.0,1.0]", "helpTip": ""}
"loss_type": "squared", # @param {"label": "loss_type", "type": "enum", "required": "true", "options": "squared,absolute", "helpTip": ""}
"max_iter": 20, # @param {"label": "max_iter", "type": "integer", "required": "true", "range": "(0,2147483647]", "helpTip": ""}
"step_size": 0.1, # @param {"label": "step_size", "type": "number", "required": "true", "range": "(0.0,none)", "helpTip": ""}
"impurity": "variance"
}
gbt_regressor____id___ = MLSGBTRegression(**params)
gbt_regressor____id___.run()
# @output {"label":"pipeline_model","name":"gbt_regressor____id___.get_outputs()['output_port_1']","type":"PipelineModel"}
