第八节 综合排序:Function Score Query 优化算分

1、算分与排序

• Elasticsearch 默认会以文档的相关度算分进⾏排序
• 可以通过指定⼀个或者多个字段进⾏排序
• 使⽤相关度算分(score)排序，不能满足某些特定条件
  • ⽆法针对相关度，对排序实现更多的控制

2、Function Score Query

2-1 Function Score Query

• 可以在查询结束后，对每⼀个匹配的⽂文档进⾏⼀系列列的重新算分，根据新⽣成的分数进行排序。

2-2 提供了⼏种默认的计算分值的函数

• Weight :为每⼀个⽂档设置⼀个简单而不被规范化的权重
• Field Value Factor:使⽤用该数值来修改 _score，例例如将 “热度”和“点赞数”作为算分的参考因素
• Random Score:为每⼀个用户使⽤⼀个不同的，随机算分结果
• 衰减函数: 以某个字段的值为标准，距离某个值越近，得分越⾼
• Script Score:自定义脚本完全控制所需逻辑

DELETE blogs

PUT /blogs/_doc/1
{
  "title":   "About popularity",
  "content": "In this post we will talk about...",
  "votes":   0
}

PUT /blogs/_doc/2
{
  "title":   "About popularity",
  "content": "In this post we will talk about...",
  "votes":   100
}

PUT /blogs/_doc/3
{
  "title":   "About popularity",
  "content": "In this post we will talk about...",
  "votes":   1000000
}

3、按受欢迎度提升权重

希望能够将点赞多的 blog，放在搜索列表 对靠前的位置。同时搜索的评分，还是要作为排序的主要依据

新的算分 = ⽼的算分 * 投票数

• 投票数为0
• 投票数很大时

POST /blogs/_search
{
  "query": {
    "function_score": {
      "query": {
        "multi_match": {
          "query":    "popularity",
          "fields": [ "title", "content" ]
        }
      },
      "field_value_factor": {
        "field": "votes"
      }
    }
  }
}

• "multi_match"
• "field_value_factor"

Output :

"max_score" : 133531.39,
    "hits" : [
      {
        "_index" : "blogs",
        "_type" : "_doc",
        "_id" : "3",
        "_score" : 133531.39,
        "_source" : {
          "title" : "About popularity",
          "content" : "In this post we will talk about...",
          "votes" : 1000000
        }
      },
      {
        "_index" : "blogs",
        "_type" : "_doc",
        "_id" : "2",
        "_score" : 13.353139,
        "_source" : {
          "title" : "About popularity",
          "content" : "In this post we will talk about...",
          "votes" : 100
        }
      },
      {
        "_index" : "blogs",
        "_type" : "_doc",
        "_id" : "1",
        "_score" : 0.0,
        "_source" : {
          "title" : "About popularity",
          "content" : "In this post we will talk about...",
          "votes" : 0
        }
      }
    ]

• "_id" : "3", "_score" : 133531.39, "votes" : 1000000
• "_id" : "2", "_score" : 13.353139, "votes" : 100
• "_id" : "1", "_score" : 0.0, "votes" : 0

4、 使⽤ Modifier 平滑曲线

• 新的算分=⽼的算分*log(1+投票数)

POST /blogs/_search
{
  "query": {
    "function_score": {
      "query": {
        "multi_match": {
          "query":    "popularity",
          "fields": [ "title", "content" ]
        }
      },
      "field_value_factor": {
        "field": "votes",
        "modifier": "log1p"
      }
    }
  }
}

Output :

• "_score" : 0.8011884,
• "_score" : 0.26763982,
• "_score" : 0.0,

5、 引⼊ Factor

• 新的算分=⽼的算分*log(1+factor*投票数)

POST /blogs/_search
{
  "query": {
    "function_score": {
      "query": {
        "multi_match": {
          "query":    "popularity",
          "fields": [ "title", "content" ]
        }
      },
      "field_value_factor": {
        "field": "votes",
        "modifier": "log1p" ,
        "factor": 0.1
      }
    }
  }
}

Output :

• "_score" : 0.66765755,
• "_score" : 0.13905862,
• "_score" : 0.0,

6、 Boost Mode 和 Max Boost

• Boost Mode
  • Multiply:算分与函数值的乘积
  • Sum:算分与函数的和
  • Min / Max:算分与函数取最⼩/ 最⼤值
  • Replace:使⽤用函数值取代算分
• Max Boost 可以将算分控制在⼀个最⼤值

POST /blogs/_search
{
  "query": {
    "function_score": {
      "query": {
        "multi_match": {
          "query":    "popularity",
          "fields": [ "title", "content" ]
        }
      },
      "field_value_factor": {
        "field": "votes",
        "modifier": "log1p" ,
        "factor": 0.1
      },
      "boost_mode": "sum",
      "max_boost": 3
    }
  }
}

Output :

6-1 Max Boost 可以将算分控制在⼀个最⼤值

• "_score" : 3.1335313,
• "_score" : 1.1749241,
• "_score" : 0.13353139,

7、⼀致性随机函数

• 使⽤用场景:⽹站的⼴告需要提⾼展现率
• 具体需求:让每个⽤户能看到不同的随机排名，但是也希望同⼀个⽤户访问时，结果的相对顺序，保持⼀致 (Consistently Random)

POST /blogs/_search
{
  "query": {
    "function_score": {
      "random_score": {
        "seed": 911119
      }
    }
  }
}

Output :

1 -> 3 -> 2

POST /blogs/_search
{
  "query": {
    "function_score": {
      "random_score": {
        "seed": 888
      }
    }
  }
}

Output :

2-> 1 -> 3

8、本节知识点回顾

• 复合查询:Function Score Query
  • 提供了了多种函数，⾃定义脚本，完全控制算分
• Field Value Factor:搜索的相关度，能够结合投票的数量进⾏重算。通过⼀些参数的设定，对算分进⾏控制
• 随机函数:⽤户提供 Seed，返回⼀个随机⼀致性的排序结果