分位数离散化训练 (QuantileDiscretizerTrainBatchOp)

Java 类名:com.alibaba.alink.operator.batch.feature.QuantileDiscretizerTrainBatchOp

Python 类名:QuantileDiscretizerTrainBatchOp

功能介绍

分位点离散可以计算选定列的分位点,然后使用这些分位点进行离散化。
生成选中列对应的q-quantile,其中可以所有列指定一个,也可以每一列对应一个

编码结果

Encode ——> INDEX

预测结果为单个token的index

Encode ——> VECTOR

预测结果为稀疏向量:

1. dropLast为true,向量中非零元个数为0或者1
2. dropLast为false,向量中非零元个数必定为1
Encode ——> ASSEMBLED_VECTOR

预测结果为稀疏向量,是预测选择列中,各列预测为VECTOR时,按照选择顺序ASSEMBLE的结果。

向量维度

Encode ——> Vector

$$ vectorSize = numBuckets - dropLast(true: 1, false: 0) + (handleInvalid: keep(1), skip(0), error(0)) $$

numBuckets: 训练参数

dropLast: 预测参数

handleInvalid: 预测参数

Token index

Encode ——> Vector
1. 正常数据: 唯一的非零元为数据所在的bucket,若 dropLast为true, 最大的bucket的值会被丢掉,预测结果为全零元

2. null: 
    2.1 handleInvalid为keep: 唯一的非零元为:numBuckets - dropLast(true: 1, false: 0)
    2.2 handleInvalid为skip: null
    2.3 handleInvalid为error: 报错

参数说明

名称 中文名称 描述 类型 是否必须? 取值范围 默认值
selectedCols 选择的列名 计算列对应的列名列表 String[] 所选列类型为 [BIGDECIMAL, BIGINTEGER, BYTE, DOUBLE, FLOAT, INTEGER, LONG, SHORT]
leftOpen 是否左开右闭 左开右闭为true,左闭右开为false Boolean true
numBuckets quantile个数 quantile个数,对所有列有效。 Integer 2
numBucketsArray quantile个数 quantile个数,每一列对应数组中一个元素。 Integer[] null

代码示例

Python 代码

from pyalink.alink import *

import pandas as pd

useLocalEnv(1)

df = pd.DataFrame([
    ["a", 1, 1, 2.0, True],
    ["c", 1, 2, -3.0, True],
    ["a", 2, 2, 2.0, False],
    ["c", 0, 0, 0.0, False]
])

batchSource = BatchOperator.fromDataframe(
    df, schemaStr='f_string string, f_long long, f_int int, f_double double, f_boolean boolean')
streamSource = StreamOperator.fromDataframe(
    df, schemaStr='f_string string, f_long long, f_int int, f_double double, f_boolean boolean')

trainOp = QuantileDiscretizerTrainBatchOp()\
    .setSelectedCols(['f_double'])\
    .setNumBuckets(8)\
    .linkFrom(batchSource)
predictBatchOp = QuantileDiscretizerPredictBatchOp()\
    .setSelectedCols(['f_double'])
predictStreamOp = QuantileDiscretizerPredictStreamOp(trainOp)\
    .setSelectedCols(['f_double'])

predictBatchOp.linkFrom(trainOp, batchSource).print()
predictStreamOp.linkFrom(streamSource) .print()

StreamOperator.execute()

Java 代码

import org.apache.flink.types.Row;

import com.alibaba.alink.operator.batch.BatchOperator;
import com.alibaba.alink.operator.batch.feature.QuantileDiscretizerPredictBatchOp;
import com.alibaba.alink.operator.batch.feature.QuantileDiscretizerTrainBatchOp;
import com.alibaba.alink.operator.batch.source.MemSourceBatchOp;
import com.alibaba.alink.operator.stream.StreamOperator;
import com.alibaba.alink.operator.stream.feature.QuantileDiscretizerPredictStreamOp;
import com.alibaba.alink.operator.stream.source.MemSourceStreamOp;
import org.junit.Test;

import java.util.Arrays;
import java.util.List;

public class QuantileDiscretizerTrainBatchOpTest {
	@Test
	public void testQuantileDiscretizerTrainBatchOp() throws Exception {
		List <Row> df = Arrays.asList(
			Row.of("a", 1, 1, 2.0, true),
			Row.of("c", 1, 2, -3.0, true),
			Row.of("a", 2, 2, 2.0, false),
			Row.of("c", 0, 0, 0.0, false)
		);
		BatchOperator <?> batchSource = new MemSourceBatchOp(df,
			"f_string string, f_long int, f_int int, f_double double, f_boolean boolean");
		StreamOperator <?> streamSource = new MemSourceStreamOp(df,
			"f_string string, f_long int, f_int int, f_double double, f_boolean boolean");
		BatchOperator <?> trainOp = new QuantileDiscretizerTrainBatchOp()
			.setSelectedCols("f_double")
			.setNumBuckets(8)
			.linkFrom(batchSource);
		BatchOperator <?> predictBatchOp = new QuantileDiscretizerPredictBatchOp()
			.setSelectedCols("f_double");
		predictBatchOp.linkFrom(trainOp, batchSource).print();
		StreamOperator <?> predictStreamOp = new QuantileDiscretizerPredictStreamOp(trainOp)
			.setSelectedCols("f_double");
		predictStreamOp.linkFrom(streamSource).print();
		StreamOperator.execute();
	}
}

运行结果

f_string f_long f_int f_double f_boolean
a 1 1 2 true
c 1 2 0 true
a 2 2 2 false
c 0 0 1 false