用ElasticSearch实现基于标签的兴趣推荐

前提

操作系统：mac
ElasticSearch 7.8

推荐原理

• 文章索引中有字段tags，存储了文章有关的标签

• 每个用户都有自己的兴趣标签tags

• 兴趣推荐就是用兴趣标签去匹配文章的标签，用户的一个兴趣标签命中N篇文章，用户的多个兴趣标签命中M篇文章，M和N有交叉，即文章中有重复，重复出现次数最多的文章就是最贴近用户兴趣的。原理理解起来简单，使用ES的目的是解决快速查询和排序的问题。

创建索引

PUT rcmd

PUT /rcmd/_mapping
{
  "properties": {
    "tags": {
      "type": "keyword",
      "store": true
    },
    "update_time": {
      "type": "date",
      "store": true
    }
  }
}

两个字段：

• tags，文章的兴趣标签，keyword类型就是不需要全文检索，标签以数组的形式存放

• update_time，更新时间，这是给兴趣推荐加一个额外的排序条件，实际项目中往往是需要结合时间和匹配度来排序的

模拟数据

POST /rcmd/_doc 
{
     "tags": [
         "布料",
         "抹布",
         "裤子",
         "衣服",
         "生活"
     ],
     "update_time": "2020-06-01T00:02:11.030"
 }

 # 再插入一条，同样标签，但是时间不一样，后面例子中有妙用

POST /rcmd/_doc
 {
     "tags": [
         "布料",
         "抹布",
         "裤子",
         "衣服",
         "生活"
     ],
     "update_time": "2020-07-01T00:02:11.030"
 }


GET /rcmd/_search

POST /rcmd/_doc
{
  "tags": [
    "啤酒",
    "米酒",
    "饮料",
    "餐饮",
    "生活"
  ],
  "update_time": "2020-06-02T00:02:11.030"
}

POST /rcmd/_doc
{
     "tags": [
         "火锅",
         "自助餐",
         "外卖",
         "烧烤",
         "餐饮"
     ],
     "update_time": "2020-06-03T00:02:11.030"
 }



POST /rcmd/_doc
{
     "tags": [
         "太阳",
         "月亮",
         "大海",
         "星星",
         "自然"
     ],
     "update_time": "2020-06-01T00:02:11.030"
 }

POST /rcmd/_doc
{
     "tags": [
         "人类",
         "动物",
         "植物",
         "地球",
         "自然"
     ],
     "update_time": "2020-06-01T00:02:11.030"
 }

POST /rcmd/_doc
{
     "tags": [
         "男人",
         "女人",
         "小孩",
         "老人",
         "人类"
     ],
     "update_time": "2020-06-02T00:02:11.030"
 }

最终数据如下

固定分数查询

GET /rcmd/_search 
 {
     "query": {
         "bool": {
             "should": [
                 {
                     "constant_score": {
                         "boost": 1,
                         "filter": {
                             "match": {
                                 "tags": "生活"
                             }
                         }
                     }
                 },
                 {
                     "constant_score": {
                         "boost": 1,
                         "filter": {
                             "match": {
                                 "tags": "衣服"
                             }
                         }
                     }
                 },
                 {
                     "constant_score": {
                         "boost": 1,
                         "filter": {
                             "match": {
                                 "tags": "火锅"
                             }
                         }
                     }
                 }
             ]
         }
     }
 }

should表达式的意义是匹配“生活”、“衣服”、“火锅”三个标签中任何一个的文章都可以返回。用constant_score查询，如果某个文章涵盖标签越多分值就越高。也就是说如果某个文章标签完全涵盖了这三个标签，那么它的分值最高的。查询结果如下：

{
  "took": 2,
  "timed_out": false,
  "_shards": {
    "total": 1,
    "successful": 1,
    "skipped": 0,
    "failed": 0
  },
  "hits": {
    "total": {
      "value": 4,
      "relation": "eq"
    },
    "max_score": 2.0,
    "hits": [
      {
        "_index": "rcmd",
        "_type": "_doc",
        "_id": "brQO63MBTdXKc2eArv9A",
        "_score": 2.0,
        "_source": {
          "tags": [
            "布料",
            "抹布",
            "裤子",
            "衣服",
            "生活"
          ],
          "update_time": "2020-06-01T00:02:11.030"
        }
      },
      {
        "_index": "rcmd",
        "_type": "_doc",
        "_id": "b7QP63MBTdXKc2eAPf_Y",
        "_score": 2.0,
        "_source": {
          "tags": [
            "布料",
            "抹布",
            "裤子",
            "衣服",
            "生活"
          ],
          "update_time": "2020-07-01T00:02:11.030"
        }
      },
      {
        "_index": "rcmd",
        "_type": "_doc",
        "_id": "cLQQ63MBTdXKc2eA6_8v",
        "_score": 1.0,
        "_source": {
          "tags": [
            "啤酒",
            "米酒",
            "饮料",
            "餐饮",
            "生活"
          ],
          "update_time": "2020-06-02T00:02:11.030"
        }
      },
      {
        "_index": "rcmd",
        "_type": "_doc",
        "_id": "cbQS63MBTdXKc2eAcP-N",
        "_score": 1.0,
        "_source": {
          "tags": [
            "火锅",
            "自助餐",
            "外卖",
            "烧烤",
            "餐饮"
          ],
          "update_time": "2020-06-03T00:02:11.030"
        }
      }
    ]
  }
}

有两篇文章涵盖了其中两个标签“生活”和“衣服”，得分为2，排到了前面。这个排序基本满足了兴趣匹配的要求。

兴趣标签权值

实际的项目中往往是用户的兴趣标签的权值不一样，假设用户的兴趣标签是[“火锅”,“生活”,“衣服”]，排在越前面的权重越高，查询的时候需要给关键词设定权重，上面的查询语句所有boost都是默认值1，现在根据需求改动权值再查询。

GET /rcmd/_search 
{
     "query": {
         "bool": {
             "should": [
                 {
                     "constant_score": {
                         "boost": 1,
                         "filter": {
                             "match": {
                                 "tags": "生活"
                             }
                         }
                     }
                 },
                 {
                     "constant_score": {
                         "boost": 4,
                         "filter": {
                             "match": {
                                 "tags": "衣服"
                             }
                         }
                     }
                 },
                 {
                     "constant_score": {
                         "boost": 6,
                         "filter": {
                             "match": {
                                 "tags": "火锅"
                             }
                         }
                     }
                 }
             ]
         }
     }
 }

分别给三个词加上权重6、4、1，查询结果如下：

{
  "took": 1,
  "timed_out": false,
  "_shards": {
    "total": 1,
    "successful": 1,
    "skipped": 0,
    "failed": 0
  },
  "hits": {
    "total": {
      "value": 4,
      "relation": "eq"
    },
    "max_score": 6.0,
    "hits": [
      {
        "_index": "rcmd",
        "_type": "_doc",
        "_id": "cbQS63MBTdXKc2eAcP-N",
        "_score": 6.0,
        "_source": {
          "tags": [
            "火锅",
            "自助餐",
            "外卖",
            "烧烤",
            "餐饮"
          ],
          "update_time": "2020-06-03T00:02:11.030"
        }
      },
      {
        "_index": "rcmd",
        "_type": "_doc",
        "_id": "brQO63MBTdXKc2eArv9A",
        "_score": 5.0,
        "_source": {
          "tags": [
            "布料",
            "抹布",
            "裤子",
            "衣服",
            "生活"
          ],
          "update_time": "2020-06-01T00:02:11.030"
        }
      },
      {
        "_index": "rcmd",
        "_type": "_doc",
        "_id": "b7QP63MBTdXKc2eAPf_Y",
        "_score": 5.0,
        "_source": {
          "tags": [
            "布料",
            "抹布",
            "裤子",
            "衣服",
            "生活"
          ],
          "update_time": "2020-07-01T00:02:11.030"
        }
      },
      {
        "_index": "rcmd",
        "_type": "_doc",
        "_id": "cLQQ63MBTdXKc2eA6_8v",
        "_score": 1.0,
        "_source": {
          "tags": [
            "啤酒",
            "米酒",
            "饮料",
            "餐饮",
            "生活"
          ],
          "update_time": "2020-06-02T00:02:11.030"
        }
      }
    ]
  }
}

可以看到包含“火锅”的文章排到了第一，包含“衣服”和“生活”的文章虽然两个词都命中，但是在权值的弱化之下排到了第二第三位。

多条件排序

GET /rcmd/_search
{
  "query": {
    "function_score": {
      "query": {
        "bool": {
          "must": [
            {
              "range": {
                "update_time": {
                  "from": "2020-06-01",
                  "to": "2020-08-01"
                }
              }
            },
            {
              "bool": {
                "should": [
                  {
                    "term": {
                      "tags": {
                        "term": "火锅",
                        "boost": 2
                      }
                    }
                  },
                  {
                    "term": {
                      "tags": {
                        "term": "衣服",
                        "boost": 1
                      }
                    }
                  },
                  {
                    "term": {
                      "tags": {
                        "term": "生活",
                        "boost": 1
                      }
                    }
                  }
                ]
              }
            }
          ]
        }
      },
      "functions": [
        {
          "gauss": {
            "update_time": {
              "scale": "3d",
              "origin": "2020-07-02T00:01:00.000"
            }
          }
        }
      ]
    }
  },
  "_source": {
    "include": [
      "tags",
      "update_time"
    ]
  },
  "from": 0,
  "size": 10
}

以上是相对完整的一个查询，首先对update_time发布时间做了限制，只选择一定范围内的数据，随后是标签的匹配，多个标签匹配条件之间是"OR"的关系，标签具有不同的权重，接下来用衰减函数gauss对update_time做衰减排序，衰减函数的意义是越近越好，scale": "3d"就是以3天为一个阶梯先对数据进行排序，相同阶梯内的数据再按照标签匹配度排序。注：gauss中的origin可以不指定 最终的查询结果：

{
  "took": 2,
  "timed_out": false,
  "_shards": {
    "total": 1,
    "successful": 1,
    "skipped": 0,
    "failed": 0
  },
  "hits": {
    "total": {
      "value": 4,
      "relation": "eq"
    },
    "max_score": 3.6649413,
    "hits": [
      {
        "_index": "rcmd",
        "_type": "_doc",
        "_id": "b7QP63MBTdXKc2eAPf_Y",
        "_score": 3.6649413,
        "_source": {
          "update_time": "2020-07-01T00:02:11.030",
          "tags": [
            "布料",
            "抹布",
            "裤子",
            "衣服",
            "生活"
          ]
        }
      },
      {
        "_index": "rcmd",
        "_type": "_doc",
        "_id": "cbQS63MBTdXKc2eAcP-N",
        "_score": 4.4511746E-28,
        "_source": {
          "update_time": "2020-06-03T00:02:11.030",
          "tags": [
            "火锅",
            "自助餐",
            "外卖",
            "烧烤",
            "餐饮"
          ]
        }
      },
      {
        "_index": "rcmd",
        "_type": "_doc",
        "_id": "cLQQ63MBTdXKc2eA6_8v",
        "_score": 1.764942E-30,
        "_source": {
          "update_time": "2020-06-02T00:02:11.030",
          "tags": [
            "啤酒",
            "米酒",
            "饮料",
            "餐饮",
            "生活"
          ]
        }
      },
      {
        "_index": "rcmd",
        "_type": "_doc",
        "_id": "brQO63MBTdXKc2eArv9A",
        "_score": 2.8566082E-32,
        "_source": {
          "update_time": "2020-06-01T00:02:11.030",
          "tags": [
            "布料",
            "抹布",
            "裤子",
            "衣服",
            "生活"
          ]
        }
      }
    ]
  }
}

同样是匹配了“衣服”和“生活”的两篇文章，一篇在最前面，一篇在最后面，是因为update_time的缘故，一篇是7月1日发布的，另一篇在6月1日，不在同一时间阶梯内，日期久远的排到了后面。中间的两篇，各自匹配了一个标签，分别是“烧烤”和“生活”，两篇文章时间阶梯没有明显的区别，然而匹配“火锅”的排到了前面，是因为“火锅”的关键词加了较高的权重。至此，我们实现了按照标签匹配文章，并且结合了时间因素和匹配度评分的兴趣推荐。

后续问题

• 本文仅仅实现了推荐中信息的匹配和排序，实际的推荐系统中还有信息流中过滤已推文章的问题，取备选文章和历史文章的交集是比较耗时的运算；此外还要解决用户Feed流推拉问题。

• 在本文中没有提及用户兴趣标签的累积操作，理论上来说只要用户读了相应的文章，就根据文章的标签给用户兴趣标签累计加分，实际项目中往往需要处理兴趣标签截断统一降权，例如标签库有1万个标签，用户进行多次阅读行为之后，一个用户和1万个标签都会有关系，分值大小不同的区别而已，如果一直根据高分的兴趣标签给用户推荐文章，兴趣标签就形成了马太效应，新的兴趣标签没有机会超过累积高分的标签，兴趣推送会越来越窄，这时候就需要截取一定数量的兴趣标签，例如截取前100个，然后统一降分，给用户的新兴趣超赶机会。

• 以上例子没有在超大数据环境下测试过，还没有具体的性能指标。