Java 类名:com.alibaba.alink.operator.batch.classification.XGBoostTrainBatchOp
Python 类名:XGBoostTrainBatchOp
XGBoost 组件是在开源社区的基础上进行包装,使功能和 PAI 更兼容,更易用。
XGBoost 算法在 Boosting 算法的基础上进行了扩展和升级,具有较好的易用性和鲁棒性,被广泛用在各种机器学习生产系统和竞赛领域。
当前支持分类,回归和排序。
名称 | 中文名称 | 描述 | 类型 | 是否必须? | 取值范围 | 默认值 |
---|---|---|---|---|---|---|
labelCol | 标签列名 | 输入表中的标签列名 | String | ✓ | ||
numRound | 树的棵树 | 树的棵树 | Integer | ✓ | ||
alpha | L1 正则项 | L1 正则项 | Double | 1.0 | ||
baseScore | Base score | Base score | Double | 0.5 | ||
colSampleByLevel | 每个树列采样 | 每个树列采样 | Double | 1.0 | ||
colSampleByNode | 每个结点列采样 | 每个结点采样 | Double | 1.0 | ||
colSampleByTree | 每个树列采样 | 每个树列采样 | Double | 1.0 | ||
eta | 学习率 | 学习率 | Double | 0.3 | ||
featureCols | 特征列名数组 | 特征列名数组,默认全选 | String[] | 所选列类型为 [BIGDECIMAL, BIGINTEGER, BYTE, DOUBLE, FLOAT, INTEGER, LONG, SHORT] | null | |
gamma | 结点分裂最小损失变化 | 节点分裂最小损失变化 | Double | 0.0 | ||
growPolicy | GrowPolicy | GrowPolicy | String | “DEPTH_WISE”, “LOSS_GUIDE” | “DEPTH_WISE” | |
interactionConstraints | interaction constraints | interaction constraints | String | null | ||
lambda | L2 正则项 | L2 正则项 | Double | 1.0 | ||
maxBin | 最大结点个数 | 最大结点个数 | Integer | 256 | ||
maxDeltaStep | Delta step | Delta step | Double | 0.0 | ||
maxDepth | 最大深度 | 最大深度 | Integer | 6 | ||
maxLeaves | 最大结点个数 | 最大结点个数 | Integer | 0 | ||
minChildWeight | 结点的最小权重 | 结点的最小权重 | Double | 1.0 | ||
monotoneConstraints | monotone constraints | monotone constraints | String | null | ||
numClass | 标签类别个数 | 标签类别个数, 多分类时有效 | Integer | 0 | ||
objective | objective | objective | String | “BINARY_LOGISTIC”, “BINARY_LOGITRAW”, “BINARY_HINGE”, “MULTI_SOFTMAX”, “MULTI_SOFTPROB” | “BINARY_LOGISTIC” | |
pluginVersion | 插件版本号 | 插件版本号 | String | “1.5.1” | ||
processType | ProcessType | ProcessType | String | “DEFAULT”, “UPDATE” | “DEFAULT” | |
refreshLeaf | RefreshLeaf | RefreshLeaf | Integer | 1 | ||
runningMode | 运行模式 | XGBoost的运行模型,ICQ速度快,但使用内存多,TRIAVIAL速度略慢,但是节省内存,按照流式方式处理。由于训练数据本身在XGBoost运行时已经被缓存进内存,所以存两份和存一份数据的资源消耗和速度对比,还需要进一步的测试。 | String | “ICQ”, “TRIVIAL” | “TRIVIAL” | |
samplingMethod | 采样方法 | 采样方法 | String | “UNIFORM”, “GRADIENT_BASED” | “UNIFORM” | |
scalePosWeight | ScalePosWeight | ScalePosWeight | Double | 1.0 | ||
singlePrecisionHistogram | single precision histogram | single precision histogram | Boolean | false | ||
sketchEps | SketchEps | SketchEps | Double | 0.03 | ||
subSample | 样本采样比例 | 样本采样比例 | Double | 1.0 | ||
treeMethod | 构建树的方法 | 构建树的方法 | String | “AUTO”, “EXACT”, “APPROX”, “HIST” | “AUTO” | |
updater | Updater | Updater | String | “grow_colmaker,prune” | ||
vectorCol | 向量列名 | 向量列对应的列名,默认值是null | String | 所选列类型为 [DENSE_VECTOR, SPARSE_VECTOR, STRING, VECTOR] | null |
** 以下代码仅用于示意,可能需要修改部分代码或者配置环境后才能正常运行!**
df = pd.DataFrame([ [0, 1, 1.1, 1.0], [1, -2, 0.9, 2.0], [0, 100, -0.01, 3.0], [1, -99, 0.1, 4.0], [0, 1, 1.1, 5.0], [1, -2, 0.9, 6.0] ]) batchSource = BatchOperator.fromDataframe( df, schemaStr='y int, x1 int, x2 double, x3 double' ) streamSource = StreamOperator.fromDataframe( df, schemaStr='y int, x1 int, x2 double, x3 double' ) trainOp = XGBoostTrainBatchOp()\ .setNumRound(1)\ .setPluginVersion('1.5.1')\ .setLabelCol('y')\ .linkFrom(batchSource) predictBatchOp = XGBoostPredictBatchOp()\ .setPredictionDetailCol('pred_detail')\ .setPredictionCol('pred')\ .setPluginVersion('1.5.1') predictStreamOp = XGBoostPredictStreamOp(trainOp)\ .setPredictionDetailCol('pred_detail')\ .setPredictionCol('pred')\ .setPluginVersion('1.5.1') predictBatchOp.linkFrom(trainOp, batchSource).print() predictStreamOp.linkFrom(streamSource).print() StreamOperator.execute()
import org.apache.flink.types.Row; import com.alibaba.alink.operator.batch.BatchOperator; import com.alibaba.alink.operator.batch.classification.XGBoostPredictBatchOp; import com.alibaba.alink.operator.batch.classification.XGBoostTrainBatchOp; import com.alibaba.alink.operator.batch.source.MemSourceBatchOp; import com.alibaba.alink.operator.stream.StreamOperator; import com.alibaba.alink.operator.stream.classification.XGBoostPredictStreamOp; import com.alibaba.alink.operator.stream.source.MemSourceStreamOp; import org.junit.Test; import java.util.Arrays; import java.util.List; public class XGBoostTrainBatchOpTest { @Test public void testXGBoostTrainBatchOp() throws Exception { List <Row> data = Arrays.asList( Row.of(0, 1, 1.1, 1.0), Row.of(1, -2, 0.9, 2.0), Row.of(0, 100, -0.01, 3.0), Row.of(1, -99, 0.1, 4.0), Row.of(0, 1, 1.1, 5.0), Row.of(1, -2, 0.9, 6.0) ); BatchOperator <?> batchSource = new MemSourceBatchOp(data, "y int, x1 int, x2 double, x3 double"); StreamOperator <?> streamSource = new MemSourceStreamOp(data, "y int, x1 int, x2 double, x3 double"); BatchOperator <?> trainOp = new XGBoostTrainBatchOp() .setNumRound(1) .setPluginVersion("1.5.1") .setLabelCol("y") .linkFrom(batchSource); BatchOperator <?> predictBatchOp = new XGBoostPredictBatchOp() .setPredictionDetailCol("pred_detail") .setPredictionCol("pred") .setPluginVersion("1.5.1"); StreamOperator <?> predictStreamOp = new XGBoostPredictStreamOp(trainOp) .setPredictionDetailCol("pred_detail") .setPredictionCol("pred") .setPluginVersion("1.5.1"); predictBatchOp.linkFrom(trainOp, batchSource).print(); predictStreamOp.linkFrom(streamSource).print(); StreamOperator.execute(); } }