使用mahout fpgrowth算法求关联规则

使用mahout fpgrowth

  首先,这篇文章的内容大部分取自国外一篇博客Finding association rules with Mahout Frequent Pattern Mining,写这个出于几个原因,一 原文是英文的;二该博客貌似还被墙了,反正我是用了goagent才看到的;三 我简化了其实验内容,单纯的用数字表示item了。

  首先是实验环境

  • jdk >= 1.6
  • maven
  • hadoop (>1.0.0)
  • mahout >= 0.7

  环境搭建就不多说了,唯一注意的是mahout按照官网的指导绝对没问题,如果安装之后报错,可能是你的hadoop版本问题,换个hadoop试试,我遇到的错就是一直

Exception in thread "main" java.lang.NoClassDefFoundError:classpath。

  我用的数据是mahout官网上提供的retail.dat,使用哪个数据没关系,mahout fpgrowth的数据格式要求如下: 

  [item id1], [item id2], [item id3]
  0, 2, 6, ...
  0, 1, 6, ...
  4, 5, ...
  ...

  间隔符可以是别的,retail.dat里用的是空格,但要注意的是使用命令行时要标志。

  这里不设置MAHOUT_LOCAL,让mahout在hadoop上跑,所以先使用hadoop命令把数据放到hdfs上,在terminal输入:

  hadoop fs -put output.dat retail.dat

  然后输入如下指令运行mahout:

  mahout fpg -i output.dat -o patterns -k 10 -method mapreduce -regex '[\ ]' -s 10

  指令的含义在mahout的网站上有详细说明,简要说下,-i表示输入,-o表示输出,-k 10表示找出和某个item相关的前十个频繁项,-method mapreduce表示使用mapreduce来运行这个作业,-regex '[\ ]'表示每个transaction里用空白来间隔item的,-s 10表示只统计最少出现10次的项。

  成功运行后在patterns文件夹里会出现四个文件或者文件夹

  • fList: 记录了每个item出现的次数的序列文件
  • frequentpatterns: 记录了包含每个item的频繁项的序列文件
  • fpGrowth
  • parallelcounting

  当然这些结果是在hdfs上面的,可以使用mahout命令查看下这些输出,在终端输入  mahout seqdumper -i patterns/frequentpatterns/part-r-00000

  第一行显示了与item7671有关的前十个事务(按出现次数排序), ([7671],80) 表示item7671出现在80个事务中. ([39, 7671],57) 表示39和7671这两个item同时出现在57个事务里。关联规则可以由以下几个参数来推导:

  • supportsupp(X) 包含集合X的事务出现的频率:  supp(X) = \frac{number\_of\_transactions\_containing\_X}{number\_of\_transactions}
  • confidenceconf(X \Longrightarrow Y) 包含x的事务中含有同时包含Y的比例: conf(X \Longrightarrow Y) = \frac{supp(X \cup Y)}{supp(X)}
  • lift 用来表示X和Y的相互独立程度: lift(X \Longrightarrow Y) = \frac{supp(X \cup Y)}{supp(X) * supp(Y)}
  • conviction 也是用来衡量X和Y的独立性的,这个值越大越好: conv(X \Longrightarrow Y) = \frac{1 - supp(Y)}{1 - conf(X \Longrightarrow Y)}

  下面用程序来推导关联规则,先把hdfs上面的几个文件放到本地来,

  hadoop fs -getmerge patterns/frequentpatterns frequentpatterns.seq

  hadoop fs -get patterns/fList fList.seq

  代码是java代码,怎么建工程都行,我是用的eclipse+maven,因为这样它可以自动帮我下载所需要的mahout的包,把两个序列文件拷到工程的根目录下,代码如下

  

复制代码
package heyong;

import java.util.ArrayList;
import java.util.HashMap;
import java.util.List;
import java.util.Map;

import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.LongWritable;
import org.apache.hadoop.io.SequenceFile;
import org.apache.hadoop.io.SequenceFile.Reader;
import org.apache.hadoop.io.Text;
import org.apache.mahout.common.Pair;
import org.apache.mahout.fpm.pfpgrowth.convertors.string.TopKStringPatterns;

public class ResultReaderS {
    public static Map<Integer, Long> readFrequency(Configuration configuration, String fileName) throws Exception {
        FileSystem fs = FileSystem.get(configuration);
        Reader frequencyReader = new SequenceFile.Reader(fs,
                new Path(fileName), configuration);
        Map<Integer, Long> frequency = new HashMap<Integer, Long>();
        Text key = new Text();
        LongWritable value = new LongWritable();
        while(frequencyReader.next(key, value)) {
            frequency.put(Integer.parseInt(key.toString()), value.get());
        }
        return frequency;
    }
 
 
    public static void readFrequentPatterns(
            Configuration configuration,
            String fileName,
            int transactionCount,
            Map<Integer, Long> frequency,
            double minSupport, double minConfidence) throws Exception {
        FileSystem fs = FileSystem.get(configuration);
 
        Reader frequentPatternsReader = new SequenceFile.Reader(fs,
                new Path(fileName), configuration);
        Text key = new Text();
        TopKStringPatterns value = new TopKStringPatterns();
 
        while(frequentPatternsReader.next(key, value)) {
            long firstFrequencyItem = -1;
            String firstItemId = null;
            List<Pair<List<String>, Long>> patterns = value.getPatterns();
            int i = 0;
            for(Pair<List<String>, Long> pair: patterns) {
                List<String> itemList = pair.getFirst();
                Long occurrence = pair.getSecond();
                if (i == 0) {
                    firstFrequencyItem = occurrence;
                    firstItemId = itemList.get(0);
                } else {
                    double support = (double)occurrence / transactionCount;
                    double confidence = (double)occurrence / firstFrequencyItem;
                    if ((support > minSupport
                            && confidence > minConfidence)) {
                        List<String> listWithoutFirstItem = new ArrayList<String>();
                        for(String itemId: itemList) {
                            if (!itemId.equals(firstItemId)) {
                                
                                listWithoutFirstItem.add(itemId);
                            }
                        }
                        String firstItem = firstItemId;
                        listWithoutFirstItem.remove(firstItemId);
                        System.out.printf(
                            "%s => %s: supp=%.3f, conf=%.3f",
                            listWithoutFirstItem,
                            firstItem,
                            support,
                            confidence);
 
                        if (itemList.size() == 2) {
                            // we can easily compute the lift and the conviction for set of
                            // size 2, so do it
                            int otherItemId = -1;
                            for(String itemId: itemList) {
                                if (!itemId.equals(firstItemId)) {
                                    otherItemId = Integer.parseInt(itemId);
                                    break;
                                }
                            }
                            long otherItemOccurrence = frequency.get(otherItemId);
                            double lift = (double)occurrence / (firstFrequencyItem * otherItemOccurrence);
                            double conviction = (1.0 - (double)otherItemOccurrence / transactionCount) / (1.0 - confidence);
                            System.out.printf(
                                ", lift=%.3f, conviction=%.3f",
                                lift, conviction);
                        }
                        System.out.printf("\n");
                    }
                }
                i++;
            }
        }
        frequentPatternsReader.close();
 
    }
 
    public static void main(String args[]) throws Exception {
 
        int transactionCount = 88162;//事务总数
        String frequencyFilename = "data/fList.seq";//
        String frequentPatternsFilename = "data/frequentpatterns.seq";
        double minSupport = 0.001;//支持度
        double minConfidence = 0.3;//置信度
 
 
        Configuration configuration = new Configuration();
        Map<Integer, Long> frequency = readFrequency(configuration, frequencyFilename);
        readFrequentPatterns(configuration, frequentPatternsFilename,
                transactionCount, frequency, minSupport, minConfidence);
 
    }
}
复制代码

  程序运行得到如下的结果

[39] => 3361: supp=0.003, conf=0.565, lift=0.000, conviction=0.977
[48] => 3361: supp=0.003, conf=0.560, lift=0.000, conviction=1.186
[39, 48] => 3361: supp=0.002, conf=0.396
[48] => 337: supp=0.001, conf=0.589, lift=0.000, conviction=1.271
[39] => 337: supp=0.001, conf=0.554, lift=0.000, conviction=0.952
[48] => 338: supp=0.009, conf=0.611, lift=0.000, conviction=1.344
[39] => 338: supp=0.008, conf=0.582, lift=0.000, conviction=1.018
[39, 48] => 338: supp=0.006, conf=0.405
[48] => 340: supp=0.005, conf=0.633, lift=0.000, conviction=1.422

………………

调整支持度和置信度的值,可以增强结果的满意度。至此,完成了使用mahout fpgrowth推导规则的一次入门实验室,灵活使用这个算法,还是可以在很多地方派上用场的。

我的实践:

数据:

I1 I2 I5
I2 I4
I2 I3
I1 I2 I4
I1 I3
I2 I3
I1 I3
I1 I2 I3 I5
I1 I2 I3

调用命令:

mahout fpg -i /user/hdfs/fp-growth/in/fpg.txt -o /user/hdfs/fp-growth/out -k 50 -method mapreduce -regex '[\ ]' -s 2

打印结果:

mahout seqdumper -i /user/hdfs/fp-growth/out/frequentpatterns/part-r-00000

Input Path: /user/hdfs/fp-growth/out/frequentpatterns/part-r-00000

Key: I1: Value: ([I1],6), ([I2, I1],4), ([I1, I3],4), ([I2, I1, I5],2), ([I2, I1, I3],2)
Key: I2: Value: ([I2],7), ([I2, I3],4), ([I2, I1],4), ([I2, I1, I5],2), ([I2, I1, I3],2), ([I2, I4],2)
Key: I3: Value: ([I3],6), ([I2, I3],4), ([I1, I3],4), ([I2, I1, I3],2)
Key: I4: Value: ([I2, I4],2)
Key: I5: Value: ([I2, I1, I5],2)
Count: 5

posted @ 2014-05-26 17:42  JamesFan  阅读(450)  评论(0编辑  收藏  举报