Find file History
Pull request Compare This branch is 4 commits ahead, 899 commits behind apache:master.
Latest commit 208908c Oct 26, 2016 @nsanthapuri nsanthapuri committed with mnilsson23 Multiseg docid (#178)
* fixed target docid check in LTRRescorer  and unit test for extracting features when there are multiple segments

(cherry picked from commit dfa0e2cc3baa72cec1b6329891d14b451effbd74)

* renamed unit test file and added comments

(cherry picked from commit fcfb661574c973b8963401d58145ddaf1942b511)

* updated tests

* fixed tabs to spaces in xml file
Permalink
..
Failed to load latest commit information.
example
src Multiseg docid (#178) Oct 26, 2016
README.md Generic models for ranking (#174) Oct 18, 2016
README.txt SOLR-8542: Added Solr Learning to Rank (LTR) plugin for reranking res… Oct 7, 2016
build.xml SOLR-8542: Added Solr Learning to Rank (LTR) plugin for reranking res… Oct 7, 2016
ivy.xml SOLR-8542: Added Solr Learning to Rank (LTR) plugin for reranking res… Oct 7, 2016

README.md

Apache Solr Learning to Rank

This is the main learning to rank integrated into solr repository. Read up on learning to rank

Apache Solr Learning to Rank (LTR) provides a way for you to extract features directly inside Solr for use in training a machine learned model. You can then deploy that model to Solr and use it to rerank your top X search results.

Test the plugin with solr/example/techproducts in a few easy steps!

Solr provides some simple example of indices. In order to test the plugin with the techproducts example please follow these steps.

  1. Compile solr and the examples

    cd solr ant dist ant server

  2. Run the example to setup the index

    ./bin/solr -e techproducts

  3. Stop solr and install the plugin:

    1. Stop solr

      ./bin/solr stop

    2. Create the lib folder

      mkdir example/techproducts/solr/techproducts/lib

    3. Install the plugin in the lib folder

      cp build/contrib/ltr/solr-ltr-7.0.0-SNAPSHOT.jar example/techproducts/solr/techproducts/lib/

    4. Replace the original solrconfig with one importing all the ltr components

      cp contrib/ltr/example/solrconfig.xml example/techproducts/solr/techproducts/conf/

  4. Run the example again

    ./bin/solr -e techproducts

    Note you could also have just restarted your collection using the admin page. You can find more detailed instructions here.

  5. Deploy features and a model

    curl -XPUT 'http://localhost:8983/solr/techproducts/schema/feature-store' --data-binary "@./contrib/ltr/example/techproducts-features.json" -H 'Content-type:application/json'

    curl -XPUT 'http://localhost:8983/solr/techproducts/schema/model-store' --data-binary "@./contrib/ltr/example/techproducts-model.json" -H 'Content-type:application/json'

  6. Have fun !

    • Access to the default feature store

      http://localhost:8983/solr/techproducts/schema/feature-store/_DEFAULT_

    • Access to the model store

      http://localhost:8983/solr/techproducts/schema/model-store

    • Perform a reranking query using the model, and retrieve the features

      http://localhost:8983/solr/techproducts/query?indent=on&q=test&wt=json&rq={!ltr%20model=linear%20reRankDocs=25%20efi.user_query=%27test%27}&fl=[features],price,score,name

BONUS: Train an actual machine learning model

  1. Download and install liblinear

  2. Change contrib/ltr/example/config.json "trainingLibraryLocation" to point to the train directory where you installed liblinear.

  3. Extract features, train a reranking model, and deploy it to Solr.

    cd contrib/ltr/example

    python train_and_upload_demo_model.py -c config.json

    This script deploys your features from config.json "featuresFile" to Solr. Then it takes the relevance judged query document pairs of "userQueriesFile" and merges it with the features extracted from Solr into a training file. That file is used to train a linear model, which is then deployed to Solr for you to rerank results.

  4. Search and rerank the results using the trained model

    http://localhost:8983/solr/techproducts/query?indent=on&q=test&wt=json&rq={!ltr%20model=ExampleModel%20reRankDocs=25%20efi.user_query=%27test%27}&fl=price,score,name

Changes to solrconfig.xml

<config>
  ...

  <!-- Query parser used to rerank top docs with a provided model -->
  <queryParser name="ltr" class="org.apache.solr.search.LTRQParserPlugin" />

  <!--  Transformer that will encode the document features in the response.
  For each document the transformer will add the features as an extra field
  in the response. The name of the field will be the the name of the
  transformer enclosed between brackets (in this case [features]).
  In order to get the feature vector you will have to
  specify that you want the field (e.g., fl="*,[features])  -->

  <transformer name="features" class="org.apache.solr.response.transform.LTRFeatureLoggerTransformerFactory" />

  <query>
    ...

    <!-- Cache for storing and fetching feature vectors -->
    <cache name="QUERY_DOC_FV"
      class="solr.search.LRUCache"
      size="4096"
      initialSize="2048"
      autowarmCount="4096"
      regenerator="solr.search.NoOpRegenerator" />
  </query>

</config>

Defining Features

In the learning to rank plugin, you can define features in a feature space using standard Solr queries. As an example:

features.json
[
{ "name": "isBook",
  "class": "org.apache.solr.ltr.feature.SolrFeature",
  "params":{ "fq": ["{!terms f=category}book"] }
},
{
  "name":  "documentRecency",
  "class": "org.apache.solr.ltr.feature.SolrFeature",
  "params": {
      "q": "{!func}recip( ms(NOW,publish_date), 3.16e-11, 1, 1)"
  }
},
{
  "name":"originalScore",
  "class":"org.apache.solr.ltr.feature.OriginalScoreFeature",
  "params":{}
},
{
  "name" : "userTextTitleMatch",
  "class" : "org.apache.solr.ltr.feature.SolrFeature",
  "params" : { "q" : "{!field f=title}${user_text}" }
},
 {
   "name" : "userFromMobile",
   "class" : "org.apache.solr.ltr.feature.ValueFeature",
   "params" : { "value" : "${userFromMobile}", "required":true }
 }
]

Defines five features. Anything that is a valid Solr query can be used to define a feature.

Filter Query Features

The first feature isBook fires if the term 'book' matches the category field for the given examined document. Since in this feature q was not specified, either the score 1 (in case of a match) or the score 0 (in case of no match) will be returned.

Query Features

In the second feature (documentRecency) q was specified using a function query. In this case the score for the feature on a given document is whatever the query returns (1 for docs dated now, 1/2 for docs dated 1 year ago, 1/3 for docs dated 2 years ago, etc..) . If both an fq and q is used, documents that don't match the fq will receive a score of 0 for the documentRecency feature, all other documents will receive the score specified by the query for this feature.

Original Score Feature

The third feature (originalScore) has no parameters, and uses the OriginalScoreFeature class instead of the SolrFeature class. Its purpose is to simply return the score for the original search request against the current matching document.

External Features

Users can specify external information that can to be passed in as part of the query to the ltr ranking framework. In this case, the fourth feature (userTextPhraseMatch) will be looking for an external field called 'user_text' passed in through the request, and will fire if there is a term match for the document field 'title' from the value of the external field 'user_text'. You can provide default values for external features as well by specifying ${myField:myDefault}, similar to how you would in a Solr config. In this case, the fifth feature (userFromMobile) will be looking for an external parameter called 'userFromMobile' passed in through the request, if the ValueFeature is : required=true, it will throw an exception if the external feature is not passed required=false, it will silently ignore the feature and avoid the scoring ( at Document scoring time, the model will consider 0 as feature value) The advantage in defining a feature as not required, where possible, is to avoid wasting caching space and time in calculating the featureScore. See the Run a Rerank Query section for how to pass in external information.

Custom Features

Custom features can be created by extending from org.apache.solr.ltr.feature.Feature, however this is generally not recommended. The majority of features should be possible to create using the methods described above.

Defining Models

Currently the Learning to Rank plugin supports 2 generalized forms of models: 1. Linear Model i.e. RankSVM, Pranking and 2. Multiple Additive Trees i.e. LambdaMART, Gradient Boosted Regression Trees (GBRT)

Linear

If you'd like to introduce a bias set a constant feature to the bias value you'd like and make a weight of 1.0 for that feature.

model.json
{
    "class":"org.apache.solr.ltr.model.LinearModel",
    "name":"myModelName",
    "features":[
        { "name": "userTextTitleMatch"},
        { "name": "originalScore"},
        { "name": "isBook"}
    ],
    "params":{
        "weights": {
            "userTextTitleMatch": 1.0,
            "originalScore": 0.5,
            "isBook": 0.1
        }

    }
}

This is an example of a toy Linear model. Class specifies the class to be using to interpret the model. Name is the model identifier you will use when making request to the ltr framework. Features specifies the feature space that you want extracted when using this model. All features that appear in the model params will be used for scoring and must appear in the features list. You can add extra features to the features list that will be computed but not used in the model for scoring, which can be useful for logging. Params are the Linear parameters.

Good library for training SVM, an example of a Linear model, is (https://www.csie.ntu.edu.tw/~cjlin/liblinear/ , https://www.csie.ntu.edu.tw/~cjlin/libsvm/) . You will need to convert the libSVM model format to the format specified above.

Multiple Additive Trees

model2.json
{
    "class":"org.apache.solr.ltr.model.MultipleAdditiveTreesModel",
    "name":"multipleadditivetreesmodel",
    "features":[
        { "name": "userTextTitleMatch"},
        { "name": "originalScore"}
    ],
    "params":{
        "trees": [
            {
                "weight" : 1,
                "root": {
                    "feature": "userTextTitleMatch",
                    "threshold": 0.5,
                    "left" : {
                        "value" : -100
                    },
                    "right": {
                        "feature" : "originalScore",
                        "threshold": 10.0,
                        "left" : {
                            "value" : 50
                        },
                        "right" : {
                            "value" : 75
                        }
                    }
                }
            },
            {
                "weight" : 2,
                "root": {
                    "value" : -10
                }
            }
        ]
    }
}

This is an example of a toy Multiple Additive Trees. Class specifies the class to be using to interpret the model. Name is the model identifier you will use when making request to the ltr framework. Features specifies the feature space that you want extracted when using this model. All features that appear in the model params will be used for scoring and must appear in the features list. You can add extra features to the features list that will be computed but not used in the model for scoring, which can be useful for logging. Params are the Multiple Additive Trees specific parameters. In this case we have 2 trees, one with 3 leaf nodes and one with 1 leaf node.

A good library for training LambdaMART, an example of Multiple Additive Trees, is ( http://sourceforge.net/p/lemur/wiki/RankLib/ ). You will need to convert the RankLib model format to the format specified above.

Deploy Models and Features

To send features run

curl -XPUT 'http://localhost:8983/solr/collection1/schema/feature-store' --data-binary @/path/features.json -H 'Content-type:application/json'

To send models run

curl -XPUT 'http://localhost:8983/solr/collection1/schema/model-store' --data-binary @/path/model.json -H 'Content-type:application/json'

View Models and Features

curl -XGET 'http://localhost:8983/solr/collection1/schema/feature-store'

curl -XGET 'http://localhost:8983/solr/collection1/schema/model-store'

Run a Rerank Query

Add to your original solr query rq={!ltr model=myModelName reRankDocs=25}

The model name is the name of the model you sent to solr earlier. The number of documents you want reranked, which can be larger than the number you display, is reRankDocs.

Pass in external information for external features

Add to your original solr query rq={!ltr reRankDocs=3 model=externalmodel efi.field1='text1' efi.field2='text2'}

Where "field1" specifies the name of the customized field to be used by one or more of your features, and text1 is the information to be pass in. As an example that matches the earlier shown userTextTitleMatch feature one could do:

rq={!ltr reRankDocs=3 model=externalmodel efi.user_text='Casablanca' efi.user_intent='movie'}

Extract features

To extract features you need to use the feature vector transformer features

fl=*,score,[features]&rq={!ltr model=yourModel reRankDocs=25}

If you use [features] together with your reranking model, it will return the array of features used by your model. Otherwise you can just ask solr to produce the features without doing the reranking:

fl=*,score,[features store=yourFeatureStore format=[dense|sparse] ]

This will return the values of the features in the given store. The format of the extracted features will be based on the format parameter. The default is sparse.

Assemble training data

In order to train a learning to rank model you need training data. Training data is what "teaches" the model what the appropriate weight for each feature is. In general training data is a collection of queries with associated documents and what their ranking/score should be. As an example:

secretary of state|John Kerry|0.66|CROWDSOURCE
secretary of state|Cesar A. Perales|0.33|CROWDSOURCE
secretary of state|New York State|0.0|CROWDSOURCE
secretary of state|Colorado State University Secretary|0.0|CROWDSOURCE

microsoft ceo|Satya Nadella|1.0|CLICK_LOG
microsoft ceo|Microsoft|0.0|CLICK_LOG
microsoft ceo|State|0.0|CLICK_LOG
microsoft ceo|Secretary|0.0|CLICK_LOG

In this example the first column indicates the query, the second column indicates a unique id for that doc, the third column indicates the relative importance or relevance of that doc, and the fourth column indicates the source. There are 2 primary ways you might collect data for use with your machine learning algorithim. The first is to collect the clicks of your users given a specific query. There are many ways of preparing this data to train a model (http://www.cs.cornell.edu/people/tj/publications/joachims_etal_05a.pdf). The general idea is that if a user sees multiple documents and clicks the one lower down, that document should be scored higher than the one above it. The second way is explicitly through a crowdsourcing platform like Mechanical Turk or CrowdFlower. These platforms allow you to show human workers documents associated with a query and have them tell you what the correct ranking should be.

At this point you'll need to collect feature vectors for each query document pair. You can use the information from the Extract features section above to do this. An example script has been included in example/train_and_upload_demo_model.py.

Explanation of the core reranking logic

An LTR model is plugged into the ranking through the LTRQParserPlugin. The plugin will read from the request the model, an instance of LTRScoringModel, plus other parameters. The plugin will generate an LTRQuery, a particular ReRankQuery. It wraps the original solr query for the first pass ranking, and uses the provided model in an LTRScoringQuery to rescore and rerank the top documents. The LTRScoringQuery will take care of computing the values of all the features and then will delegate the final score generation to the LTRScoringModel.

Speeding up the weight creation with threads

About half the time for ranking is spent in the creation of weights for each feature used in ranking. If the number of features is significantly high (say, 500 or more), this increases the ranking overhead proportionally. To alleviate this problem, parallel weight creation is provided as a configurable option. In order to use this feature, the following lines need to be added to the solrconfig.xml

<config>
  <!-- Query parser used to rerank top docs with a provided model -->
  <queryParser name="ltr" class="org.apache.solr.search.LTRQParserPlugin">
     <int name="threadModule.totalPoolThreads">10</int> <!-- Maximum threads to share for all requests -->
     <int name="threadModule.numThreadsPerRequest">5</int> <!-- Maximum threads to use for a single requests-->
  </queryParser>

  <!-- Transformer for extracting features -->
  <transformer name="features" class="org.apache.solr.response.transform.LTRFeatureLoggerTransformerFactory">
     <int name="threadModule.totalPoolThreads">10</int> <!-- Maximum threads to share for all requests -->
     <int name="threadModule.numThreadsPerRequest">5</int> <!-- Maximum threads to use for a single requests-->
  </transformer>
</config>

The threadModule.totalPoolThreads option limits the total number of threads to be used across all query instances at any given time. threadModule.numThreadsPerRequest limits the number of threads used to process a single query. In the above example, 10 threads will be used to services all queries and a maximum of 5 threads to service a single query. If the solr instances is expected to receive no more than one query at a time, it is best to set both these numbers to the same value. If multiple queries need to serviced simultaneously, the numbers can be adjusted based on the expected response times. If the value of threadModule.numThreadsPerRequest is higher, the reponse time for a single query will be improved upto a point. If multiple queries are serviced simultaneously, the threadModule.totalPoolThreads imposes a contention between the queries if (threadModule.numThreadsPerRequest*total parallel queries > threadModule.totalPoolThreads).