Ranking Evaluation API

The ranking evaluation API allows to evaluate the quality of ranked search results over a set of typical search queries. Given this set of queries and a list of manually rated documents, the _rank_eval endpoint calculates and returns typical information retrieval metrics like mean reciprocal rank, precision or discounted cumulative gain.

Search quality evaluation starts with looking at the users of your search application, and the things that they are searching for. Users have a specific information need, e.g. they are looking for gift in a web shop or want to book a flight for their next holiday. They usually enter some search terms into a search box or some other web form. All of this information, together with meta information about the user (e.g. the browser, location, earlier preferences etc…​) then gets translated into a query to the underlying search system.

The challenge for search engineers is to tweak this translation process from user entries to a concrete query in such a way, that the search results contain the most relevant information with respect to the users information need. This can only be done if the search result quality is evaluated constantly across a representative test suite of typical user queries, so that improvements in the rankings for one particular query doesn’t negatively effect the ranking for other types of queries.

In order to get started with search quality evaluation, three basic things are needed:

The ranking evaluation API provides a convenient way to use this information in a ranking evaluation request to calculate different search evaluation metrics. This gives a first estimation of your overall search quality and give you a measurement to optimize against when fine-tuning various aspect of the query generation in your application.

In its most basic form, a request to the _rank_eval endpoint has two sections:

GET /my_index/_rank_eval
{
    "requests": [ ... ], 
    "metric": { 
      "mean_reciprocal_rank": { ... } 
   }
}

The request section contains several search requests typical to your application, along with the document ratings for each particular search request, e.g.

    "requests": [
        {
            "id": "amsterdam_query", 
            "request": { 
                "query": { "match": { "text": "amsterdam" }}
            },
            "ratings": [ 
                 { "_index": "my_index", "_id": "doc1", "rating": 0 },
                 { "_index": "my_index", "_id": "doc2", "rating": 3},
                 { "_index": "my_index", "_id": "doc3", "rating": 1 }
            ]
        },
        {
            "id": "berlin_query",
            "request": {
                "query": { "match": { "text": "berlin" }}
            },
            "ratings": [
                { "_index": "my_index", "_id": "doc1", "rating": 1 }
            ]
        }
    ]

A document rating can be any integer value that expresses the relevance of the document on a user defined scale. For some of the metrics, just giving a binary rating (e.g. 0 for irrelevant and 1 for relevant) will be sufficient, other metrics can use a more fine grained scale.

As an alternative to having to provide a single query per test request, it is possible to specify query templates in the evaluation request and later refer to them. Queries with similar structure that only differ in their parameters don’t have to be repeated all the time in the requests section this way. In typical search systems where user inputs usually get filled into a small set of query templates, this helps making the evaluation request more succinct.

GET /my_index/_rank_eval
{
   [...]
  "templates": [
     {
        "id": "match_one_field_query",  
        "template": { 
            "inline": {
                "query": {
                  "match": { "{{field}}": { "query": "{{query_string}}" }}
                }
            }
        }
     }
  ],
  "requests": [
      {
         "id": "amsterdam_query"
         "ratings": [ ... ],
         "template_id": "match_one_field_query", 
         "params": { 
            "query_string": "amsterdam",
            "field": "text"
          }
     },
    [...]
}

The metric section determines which of the available evaluation metrics is going to be used. Currently, the following metrics are supported:

This metric measures the number of relevant results in the top k search results. Its a form of the well known Precision metric that only looks at the top k documents. It is the fraction of relevant documents in those first k search. A precision at 10 (P@10) value of 0.6 then means six out of the 10 top hits are relevant with respect to the users information need.

P@k works well as a simple evaluation metric that has the benefit of being easy to understand and explain. Documents in the collection need to be rated either as relevant or irrelevant with respect to the current query. P@k does not take into account where in the top k results the relevant documents occur, so a ranking of ten results that contains one relevant result in position 10 is equally good as a ranking of ten results that contains one relevant result in position 1.

GET /twitter/_rank_eval
{
    "requests": [
    {
        "id": "JFK query",
        "request": { "query": { "match_all": {}}},
        "ratings": []
    }],
    "metric": {
      "precision": {
        "k" : 20,
        "relevant_rating_threshold": 1,
        "ignore_unlabeled": false
      }
   }
}

The precision metric takes the following optional parameters

For every query in the test suite, this metric calculates the reciprocal of the rank of the first relevant document. For example finding the first relevant result in position 3 means the reciprocal rank is 1/3. The reciprocal rank for each query is averaged across all queries in the test suite to give the mean reciprocal rank.

GET /twitter/_rank_eval
{
    "requests": [
    {
        "id": "JFK query",
        "request": { "query": { "match_all": {}}},
        "ratings": []
    }],
    "metric": {
        "mean_reciprocal_rank": {
            "k" : 20,
            "relevant_rating_threshold" : 1
        }
    }
}

The mean_reciprocal_rank metric takes the following optional parameters

In contrast to the two metrics above, discounted cumulative gain takes both, the rank and the rating of the search results, into account.

The assumption is that highly relevant documents are more useful for the user when appearing at the top of the result list. Therefore, the DCG formula reduces the contribution that high ratings for documents on lower search ranks have on the overall DCG metric.

GET /twitter/_rank_eval
{
    "requests": [
    {
        "id": "JFK query",
        "request": { "query": { "match_all": {}}},
        "ratings": []
    }],
    "metric": {
       "dcg": {
            "k" : 20,
            "normalize": false
       }
    }
}

The dcg metric takes the following optional parameters:

Expected Reciprocal Rank (ERR) is an extension of the classical reciprocal rank for the graded relevance case (Olivier Chapelle, Donald Metzler, Ya Zhang, and Pierre Grinspan. 2009. Expected reciprocal rank for graded relevance.)

It is based on the assumption of a cascade model of search, in which a user scans through ranked search results in order and stops at the first document that satisfies the information need. For this reason, it is a good metric for question answering and navigation queries, but less so for survey oriented information needs where the user is interested in finding many relevant documents in the top k results.

The metric models the expectation of the reciprocal of the position at which a user stops reading through the result list. This means that relevant document in top ranking positions will contribute much to the overall score. However, the same document will contribute much less to the score if it appears in a lower rank, even more so if there are some relevant (but maybe less relevant) documents preceding it. In this way, the ERR metric discounts documents which are shown after very relevant documents. This introduces a notion of dependency in the ordering of relevant documents that e.g. Precision or DCG don’t account for.

GET /twitter/_rank_eval
{
    "requests": [
    {
        "id": "JFK query",
        "request": { "query": { "match_all": {}}},
        "ratings": []
    }],
    "metric": {
       "expected_reciprocal_rank": {
            "maximum_relevance" : 3,
            "k" : 20
       }
    }
}

The expected_reciprocal_rank metric takes the following parameters:

The response of the _rank_eval endpoint contains the overall calculated result for the defined quality metric, a details section with a breakdown of results for each query in the test suite and an optional failures section that shows potential errors of individual queries. The response has the following format:

{
    "rank_eval": {
        "metric_score": 0.4, 
        "details": {
            "my_query_id1": { 
                "metric_score": 0.6, 
                "unrated_docs": [ 
                    {
                        "_index": "my_index",
                        "_id": "1960795"
                    }, [...]
                ],
                "hits": [
                    {
                        "hit": { 
                            "_index": "my_index",
                            "_type": "page",
                            "_id": "1528558",
                            "_score": 7.0556192
                        },
                        "rating": 1
                    }, [...]
                ],
                "metric_details": { 
                    "precision" : {
                        "relevant_docs_retrieved": 6,
                        "docs_retrieved": 10
                    }
                }
            },
            "my_query_id2" : { [...] }
        },
        "failures": { [...] }
    }
}