用 Mahout 和 Elasticsearch 实现推荐系统

原文地址

本文内容

  • 软件
  • 步骤
  • 控制相关性
  • 总结
  • 参考资料

本文介绍如何用带 Apache Mahout 的 MapR Sandbox for Hadoop 和 Elasticsearch 搭建推荐引擎,只需要很少的代码。

This tutorial will give step-by-step instructions on how to:

  • 使用的电影评分数据位于 http://grouplens.org/datasets/movielens/
  • 使用 Apache Mahout 的协同过滤(collaborative filtering)搭建和训练机器学习模型
  • 使用 Elasticsearch 的搜索技术简化推荐系统的开发

软件

该文章运行在 MapReduce Sandbox。还要求在 Sandbox 上安装 Elasticsearch 和 Mahout。

  • 从 http://grouplens.org/datasets/movielens/ 下载 10M MovieLens 数据
  • 安装 Mahout
  • 安装 Elasticsearch

步骤

Step 1: 索引(Index)电影元数据到 Elasticsearch


在 Elasticsearch 中,默认情况下,文档的所有字段都会被索引。最简单的文档是只有一级 JSON 结构。文档包含在索引中,文档中的类型告诉 Elasticsearch 如何解释文档中的字段。

你可以把 Elasticsearch 的索引看做是关系型数据库中的数据库实例,而类型看做是数据库表,字段看做表定义(但是这个字段,在 Elasticsearch 中的意义更广泛),文档看做是表的某行记录。

针对本例,文档类型是 film。并具有如下字段:电影ID(id)、标题(title)、上映时间(year)、电影类型/标签(genre,基因)、指示(indicators)、indicators数组的数量(numFields):

{
 "id": "65006",
 "title": "Impulse",
 "year": "2008",
 "genre": ["Mystery","Thriller"],
 "indicators": ["154","272",”154","308", "535", "583", "593", "668", "670", "680", "702", "745"],
 "numFields": 12
}

通过 9200 端口访问 Elasticsearch RESTful API 与其通信,或者命令行用 curl 命令。参看 Elasticsearch REST interface 和 Elasticsearch 101 tutorial。

curl -X<VERB> 'http://<HOST>/<PATH>?<QUERY_STRING>' -d '<BODY>'

使用 Elasticsearch's REST API 的 put mapping 命令可以定义文档的类型。下面的请求在 bigmovie 索引中创建名为 film 的映射(mapping)。该映射定义一个类型为 integer 类型的 numFields 字段。默认情况,所有字段都被存储并索引,整型也如此。

curl -XPUT 'http://localhost:9200/bigmovie' -d '
{
  "mappings": {
    "film" : {
      "properties" : {
        "numFields" : { "type" :   "integer" }
      }
    }
  }
}'
电影信息包含在 movies.dat 文件中。文件的每行表示一部电影,字段的含义如下所示:
MovieID::Title::Genres

例如:

65006::Impulse (2008)::Mystery|Thriller

用 Mahout 和 Elasticsearch 实现推荐系统_第1张图片

图 1 电影《冲动(Impulse)》(2008)、类型“悬疑/惊悚”

下面 Python 脚本把 movies.dat 文件中的数据转换成 JSON 格式,以便导入 Elasticsearch:

import re
import json
count=0
with open('movies.dat','rb') as csv_file:
   content = csv_file.readlines()
   for line in content:
        fixed = re.sub("::", "\t", line).rstrip().split("\t")
   if len(fixed)==3:
          title = re.sub(" \(.*\)$", "", re.sub('"','', fixed[1]))
          genre = fixed[2].split('|')
          print '{ "create" : { "_index" : "bigmovie", "_type" : "film",
          "_id" : "%s" } }' %  fixed[0]
          print '{ "id": "%s", "title" : "%s", "year":"%s" , "genre":%s }'
          % (fixed[0],title, fixed[1][-5:-1], json.dumps(genre))

运行该 Python 文件,转换结果输出到 index.json:

$ python index.py > index.json

将产生如下 Elasticsearch 需要的格式:

{ "create" : { "_index" : "bigmovie", "_type" : "film", "_id" : "1" } }
{ "id": "1", "title" : "Toy Story", "year":"1995" , "genre":["Adventure", "Animation", "Children", "Comedy", "Fantasy"] }
{ "create" : { "_index" : "bigmovie", "_type" : "film", "_id" : "2" } }
{ "id": "2", "title" : "Jumanji", "year":"1995" , "genre":["Adventure", "Children", "Fantasy"] }

文件中的每行创建索引和类型,并添加电影信息。这是利用 Elasticsearch 批量导入数据。

Elasticsearch 批量 API 可以执行对索引的操作,用同一个 API,不同的 http 请求(如 get、put、post、delete)。下面命令让 Elasticsearch 批量加载 index.json 文中的内容:

curl -s -XPOST localhost:9200/_bulk --data-binary @index.json; echo

加载电影信息后,你就可以利用 REST API 进行查询了。你也可以使用 Chrome 的 Elasticsearch 插件——Sense 进行操作(Kibana 4 提供的一个插件)。示例如下所示:

下面是检索 id 为 1237的电影:

Step 2: 使用 Mahout 从用户评分数据中创建 Movie indicators


评分包含在 ratings.dat 文件中。该文件每行表示某个用户对某个电影的评分,格式如下所示:

UserID::MovieID::Rating::Timestamp

例如:

71567::2294::5::912577968
71567::2338::2::912578016

ratings.data 文件用 "::" 做分隔符,转换成 tab 后 Mahout 才能使用。可以用 sed 命令把 :: 替换成 tab:

sed -i 's/::/\t/g' ratings.dat

该命令打开文件,把"::" 替换成"\t" 后,重新保存。Updates are only supported with MapR NFS and thus this command probably won't work on other NFS-on-Hadoop implementations. MapR Direct Access NFS allows files to be modified (supports random reads and writes) and accessed via mounting the Hadoop cluster over NFS.

sed 命令会产生如下格式的内容,该格式可以作为 Mahout 的输入:

71567    2294    5    912580553
71567    2338    2    912580553

一般格式为:item1 item2 rating timestamp,即“物品1 物品2 评分”,本例不使用 timestamp。

启动 Mahout 物品相似度(itemsimilarity)作业,命令如下所示:

 mahout itemsimilarity \
  --input /user/user01/mlinput/ratings.dat \
  --output /user/user01/mloutput \
  --similarityClassname SIMILARITY_LOGLIKELIHOOD \
  --booleanData TRUE \
  --tempDir /user/user01/temp

The argument “-s SIMILARITY_LOGLIKELIHOOD” tells the recommender to use the Log Likelihood Ratio (LLR) method for determining which items co-occur anomalously often and thus which co-occurrences can be used as indicators of preference. 相似度默认是 0.9;this can be adjusted based on the use case with the --threshold parameter, which will discard pairs with lower similarity (the default is a fine choice). Mahout 通过启动很多 Hadoop MapReduce 作业计算推荐,最后将产生输出文件,该文件位于 /user/user01/mloutput 目录。输出文件格式如下所示

64957   64997   0.9604835425701245
64957   65126   0.919355104432831
64957   65133   0.9580439772229588

一般格式为:item1id item2id similarity,即“物品1 物品2 相似度”。

Step 3: 添加 Movie indicators 到 Elasticsearch 的电影文档


下一步,我们从上面的输出文件添加 indicators 到 Elasticsearch 的 film 文档。例如,把电影的 indicators 放到 indicators 字段:

{
  "id": "65006",
  "title": "Impulse",
  "year": "2008",
  "genre": ["Mystery","Thriller"],
  "indicators": ["1076", "1936", "2057", "2204"],
  "numFields": 4
}

左面的表显示文档中包含 indicator 的内容,右边的表显示哪些文档包含某个 indicator:

图 2 文档与 indicator

如果想要检索 indicator 为 1237 551 的电影,那么本例将返回 id 为 8298 的文档(电影)。如果检索 1237 551,那么将返回 id 为 8298、3 和 64418 的电影。

下面脚本将读取 Mahout 的输出文件 part-r-00000,为每部电影创建 indicator 数组,然后输出 JSON 文件,用该文件更新 Elasticsearch bigmovie 索引的 film 类型的 indicator 字段。

import fileinput
from string import join
import json
import csv
import json
### read the output from MAHOUT and collect into hash ###
with open('/user/user01/mloutput/part-r-00000','rb') as csv_file:
    csv_reader = csv.reader(csv_file,delimiter='\t')
    old_id = ""
    indicators = []
    update = {"update" : {"_id":""}}
    doc = {"doc" : {"indicators":[], "numFields":0}}
    for row in csv_reader:
        id = row[0]
        if (id != old_id and old_id != ""):
            update["update"]["_id"] = old_id
            doc["doc"]["indicators"] = indicators
            doc["doc"]["numFields"] = len(indicators)
            print(json.dumps(update))
            print(json.dumps(doc))
            indicators = [row[1]]
        else:
            indicators.append(row[1])
        old_id = id

下面命令会执行 update.py 的 Python 脚本,并输出 update.json:

$ python update.py > update.json

上面 Python 脚本将创建如下内容的文件:

{"update": {"_id": "1"}}
{"doc": {"indicators": ["75", "118", "494", "512", "609", "626", "631", "634", "648", "711", "761", "810", "837", "881", "910", "1022", "1030", "1064", "1301", "1373", "1390", "1588", "1806", "2053", "2083", "2090", "2096", "2102", "2286", "2375", "2378", "2641", "2857", "2947", "3147", "3429", "3438", "3440", "3471", "3483", "3712", "3799", "3836", "4016", "4149", "4544", "4545", "4720", "4732", "4901", "5004", "5159", "5309", "5313", "5323", "5419", "5574", "5803", "5841", "5902", "5940", "6156", "6208", "6250", "6383", "6618", "6713", "6889", "6890", "6909", "6944", "7046", "7099", "7281", "7367", "7374", "7439", "7451", "7980", "8387", "8666", "8780", "8819", "8875", "8974", "9009", "25947", "27721", "31660", "32300", "33646", "40339", "42725", "45517", "46322", "46559", "46972", "47384", "48150", "49272", "55668", "63808"], "numFields": 102}}
{"update": {"_id": "2"}}
{"doc": {"indicators": ["15", "62", "153", "163", "181", "231", "239", "280", "333", "355", "374", "436", "473", "485", "489", "502", "505", "544", "546", "742", "829", "1021", "1474", "1562", "1588", "1590", "1713", "1920", "1967", "2002", "2012", "2045", "2115", "2116", "2139", "2143", "2162", "2296", "2338", "2399", "2408", "2447", "2616", "2793", "2798", "2822", "3157", "3243", "3327", "3438", "3440", "3477", "3591", "3614", "3668", "3802", "3869", "3968", "3972", "4090", "4103", "4247", "4370", "4467", "4677", "4686", "4846", "4967", "4980", "5283", "5313", "5810", "5843", "5970", "6095", "6383", "6385", "6550", "6764", "6863", "6881", "6888", "6952", "7317", "8424", "8536", "8633", "8641", "26870", "27772", "31658", "32954", "33004", "34334", "34437", "39419", "40278", "42011", "45210", "45447", "45720", "48142", "50347", "53464", "55553", "57528"], "numFields": 106}}

在命令行,用 curl 命令调用 Elasticsearch REST bulk 请求,把该文件 update.json 作为输入,就可以更新 indicator 字段:

$ curl -s -XPOST localhost:9200/bigmovie/film/_bulk --data-binary @update.json; echo

Step 4: 检索 Film 索引的 indicator 字段进行推荐


现在,你就可以检索 film 的 indicator 字段进行查询并推荐。例如,某人喜欢电影 1237 和 551,你想推荐类似的电影,可以执行如下 Elasticsearch 查询获得推荐,将返回indicator 数组为 1237 和 551 的电影,即 1237=Seventh Seal(第七封印),551=Nightmare Before Christmas(圣诞夜惊魂)

curl 'http://localhost:9200/bigmovie/film/_search?pretty' -d '
{
  "query": {
    "function_score": {
      "query": {
         "bool": {
           "must": [ { "match": { "indicators":"1237 551"} } ],
           "must_not": [ { "ids": { "values": ["1237", "551"] } } ]
         }
      },
      "functions":[ {"random_score": {"seed":"48" } } ],
      "score_mode":"sum"
    }
  },
  "fields":["_id","title","genre"],
  "size":"8"
}'

上面查询 indicator 为 1237 或 551,并且不是 1237 或 551 的电影。下面示例使用 Sense 插件进行查询,右边是检索结果,推荐结果是 “A Man Named Pearl(这个是纪录片)” 和 “Used People(寡妇三弄)”。

控制相关性


全文检索引擎根据相关度排序,Elasticsearch 用 _score 字段表示文档的相关度分数(relevance score)。function_score 允许你查询时修改该分数。random_score 用一个种子变量使用散列生成分数。Elasticsearch 查询如下所示,random_score 函数用于把变量添加到检索结果,以便完成 dithering:

  "query": {
    "function_score": {
      "query": {
         "bool": {
           "must": [ { "match": { "indicators":"1237 551"} } ],
           "must_not": [ { "ids": { "values": ["1237", "551"] } } ]
         }
      },
      "functions":[ {"random_score": {"seed":"48" } } ],
      "score_mode":"sum"
    }
  }

相关性抖动(dithering)有意地包含排名靠,但相关性较低的结果,以便拓展训练数据,提供给推荐引擎。如果没有 dithering,那么明天的训练数据仅仅是教模型今天已经知道的事情。增加 dithering, 会帮助拓展推荐模型。如果模型给出的答案接近优秀的,那么 dithering 可以帮助找到正确答案。有效的 dithering 会减少今天的准确性,而改进明天的训练数据(和未来的性能,算法的准确性也属于性能的范畴),换句话说,为了让将来的推荐准确,需要减少过去对将来的影响。

总结


We showed in this tutorial how to use Apache Mahout and Elasticsearch with the MapR Sandbox to build a basic recommendation engine. You can go beyond a basic recommender and get even better results with a few simple additions to the design to add cross recommendation of items, which leverages a variety of interactions and items for making recommendations. You can find more information about these technologies here:

参考资料


若想学习更多关于推荐引擎的组件和逻辑,参看 "An Inside Look at the Components of a Recommendation Engine",该文章详细描述了推荐引擎的架构、Mahout 协同过滤(collaborative filtering)和 Elasticsearch 检索引擎。

更多关于推荐引擎、机器学习和 Elasticsearch 的资源,如下所示:

  • 实践机器学习: Innovations in Recommendations
  • An Invitation to Practical Machine Learning
  • 构建一个简单的推荐器
  • Jump-Start Your Recommendation Engine on Hadoop
  • MapR 快速解决方案 - 推荐引擎演示
  • Elasticsearch: The Definitive Guide
  • Mahout 基于物品推荐的介绍

Tutorial Category Reference:

  • ElasticSearch
  • Mahout

你可能感兴趣的:(用 Mahout 和 Elasticsearch 实现推荐系统)