字符串最近邻训练 (StringNearestNeighborTrainBatchOp)

Java 类名:com.alibaba.alink.operator.batch.similarity.StringNearestNeighborTrainBatchOp

Python 类名:StringNearestNeighborTrainBatchOp

功能介绍

本算法支持Levenshtein Distance,Longest Common SubString,String Subsequence Kernel,Cosine四种相似度精确计算方式,通过选择metric参数可计算不同的相似度。

该功能由训练和预测组成,支持计算1. 求最近邻topN 2. 求radius范围内的邻居。该功能由预测时候的topN和radius参数控制, 如果填写了topN,则输出最近邻,如果填写了radius,则输出radius范围内的邻居。

Levenshtein(Levenshtein Distance), 相似度=(1-距离)/length,length为两个字符长度的最大值,应选metric的参数为LEVENSHTEIN_SIM。

LCS(Longest Common SubString), 相似度=(1-距离)/length,length为两个字符长度的最大值,应选择metric的参数为LCS_SIM。

SSK(String Subsequence Kernel)支持相似度计算,应选择metric的参数为SSK。

Cosine(Cosine)支持相似度计算,应选择metric的参数为COSINE。

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
idCol id列名 id列名 String
selectedCol 选中的列名 计算列对应的列名 String 所选列类型为 [STRING]
lambda 匹配字符权重 匹配字符权重,SSK中使用 Double 0.5
metric 距离类型 用于计算的距离类型 String “LEVENSHTEIN_SIM”, “LEVENSHTEIN”, “LCS_SIM”, “LCS”, “SSK”, “COSINE” “LEVENSHTEIN_SIM”
windowSize 窗口大小 窗口大小 Integer x >= 1 2

代码示例

Python 代码

from pyalink.alink import *

import pandas as pd

useLocalEnv(1)

df = pd.DataFrame([
    [0, "abcde", "aabce"],
    [1, "aacedw", "aabbed"],
    [2, "cdefa", "bbcefa"],
    [3, "bdefh", "ddeac"],
    [4, "acedm", "aeefbc"]
])

inOp = BatchOperator.fromDataframe(df, schemaStr='id long, text1 string, text2 string')

train = StringNearestNeighborTrainBatchOp().setIdCol("id").setSelectedCol("text1").setMetric("LEVENSHTEIN_SIM").linkFrom(inOp)
predict = StringNearestNeighborPredictBatchOp().setSelectedCol("text2").setTopN(3).linkFrom(train, inOp)
predict.print()

Java 代码

import org.apache.flink.types.Row;

import com.alibaba.alink.operator.batch.BatchOperator;
import com.alibaba.alink.operator.batch.similarity.StringNearestNeighborPredictBatchOp;
import com.alibaba.alink.operator.batch.similarity.StringNearestNeighborTrainBatchOp;
import com.alibaba.alink.operator.batch.source.MemSourceBatchOp;
import org.junit.Test;

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

public class StringNearestNeighborTrainBatchOpTest {
	@Test
	public void testStringNearestNeighborTrainBatchOp() throws Exception {
		List <Row> df = Arrays.asList(
			Row.of(0, "abcde", "aabce"),
			Row.of(1, "aacedw", "aabbed"),
			Row.of(2, "cdefa", "bbcefa"),
			Row.of(3, "bdefh", "ddeac"),
			Row.of(4, "acedm", "aeefbc")
		);
		BatchOperator <?> inOp = new MemSourceBatchOp(df, "id int, text1 string, text2 string");
		BatchOperator <?> train = new StringNearestNeighborTrainBatchOp().setIdCol("id").setSelectedCol("text1")
			.setMetric("LEVENSHTEIN_SIM").linkFrom(inOp);
		BatchOperator <?> predict = new StringNearestNeighborPredictBatchOp().setSelectedCol("text2").setTopN(3)
			.linkFrom(train, inOp);
		predict.print();
	}
}

运行结果

id text1 text2
0 abcde {“ID”:“[0,1,4]”,“METRIC”:“[0.6,0.5,0.19999999999999996]”}
1 aacedw {“ID”:“[1,0,4]”,“METRIC”:“[0.5,0.33333333333333337,0.33333333333333337]”}
2 cdefa {“ID”:“[2,3,1]”,“METRIC”:“[0.5,0.5,0.33333333333333337]”}
3 bdefh {“ID”:“[2,3,4]”,“METRIC”:“[0.4,0.4,0.19999999999999996]”}
4 acedm {“ID”:“[4,3,2]”,“METRIC”:“[0.33333333333333337,0.33333333333333337,0.33333333333333337]”}