Building a Real-time, Solr-powered Recommendation Engine
Trey GraingerManager, Search Technology Development
@
Lucene Revolution 2012 - Boston
Overview
• Overview of Search & Matching Concepts• Recommendation Approaches in Solr:• Attribute-based• Hierarchical Classification• Concept-based• More-like-this• Collaborative Filtering• Hybrid Approaches
• Important Considerations & Advanced Capabilities @ CareerBuilder
My Background
Trey Grainger• Manager, Search Technology Development@ CareerBuilder.com
Relevant Background• Search & Recommendations• High-volume, N-tier Architectures• NLP, Relevancy Tuning, user group testing, & machine learning
Fun Side Projects• Founder and Chief Engineer @ .com
• Currently co-authoring Solr in Action book… keep your eyes out for the early access release from Manning Publications
About Search @CareerBuilder
• Over 1 million new jobs each month • Over 45 million actively searchable resumes• ~250 globally distributed search servers (in the
U.S., Europe, & Asia) • Thousands of unique, dynamically generated
indexes• Hundreds of millions of search documents• Over 1 million searches an hour
Search Products @
Redefining “Search Engine”
• “Lucene is a high-performance, full-featured text search engine library…”
Yes, but really…
• Lucene is a high-performance, fully-featured token matching and scoring library… which can perform full-text searching.
Redefining “Search Engine”
or, in machine learning speak:
• A Lucene index is a multi-dimensional sparse matrix… with very fast and powerful lookup capabilities.
• Think of each field as a matrix containing each term mapped to each document
The Lucene Inverted Index (traditional text example)
Term Documentsa doc1 [2x]
brown doc3 [1x] , doc5 [1x]
cat doc4 [1x]
cow doc2 [1x] , doc5 [1x]
… ...
once doc1 [1x], doc5 [1x]
over doc2 [1x], doc3 [1x]
the doc2 [2x], doc3 [2x], doc4[2x], doc5 [1x]
… …
Document Content Fielddoc1 once upon a time, in a land
far, far awaydoc2 the cow jumped over the
moon.doc3 the quick brown fox
jumped over the lazy dog.doc4 the cat in the hatdoc5 The brown cow said “moo”
once.… …
What you SEND to Lucene/Solr:How the content is INDEXED into Lucene/Solr (conceptually):
Match Text Queries to Text Fields
/solr/select/?q=jobcontent: (software engineer)
Job Content Field Documents… …engineer doc1, doc3, doc4,
doc5…mechanical doc2, doc4, doc6… …software doc1, doc3, doc4,
doc7, doc8… …
doc5
doc7 doc8
doc1 doc3 doc4
engineer
software
software engineer
• Lucene/Solr is a text search matching engine
• When Lucene/Solr search text, they are matching tokens in the query with tokens in index
• Anything that can be searched upon can form the basis of matching and scoring:– text, attributes, locations, results of functions, user
behavior, classifications, etc.
Beyond Text Searching
Business Case for Recommendations
• For companies like CareerBuilder, recommendations can provide as much or even greater business value (i.e. views, sales, job applications) than user-driven search capabilities.
• Recommendations create stickiness to pull users back to your company’s website, app, etc.
• What are recommendations?… searches of relevant content for a user
Approaches to Recommendations• Content-based
– Attribute based• i.e. income level, hobbies, location, experience
– Hierarchical• i.e. “medical//nursing//oncology”, “animal//dog//terrier”
– Textual Similarity• i.e. Solr’s MoreLikeThis Request Handler & Search Handler
– Concept Based• i.e. Solr => “software engineer”, “java”, “search”, “open source”
• Behavioral Based • Collaborative Filtering: “Users who liked that also liked this…”
• Hybrid Approaches
Content-based Recommendation Approaches
Attribute-based Recommendations• Example: Match User Attributes to Item Attribute Fields
Janes_Profile:{Industry:”healthcare”, Locations:”Boston, MA”, JobTitle:”Nurse Educator”, Salary:{ min:40000, max:60000 },
}
/solr/select/?q=(jobtitle:”nurse educator”^25 OR jobtitle:(nurse educator)^10) AND ((city:”Boston” AND state:”MA”)^15 OR state:”MA”) AND _val_:”map(salary,40000,60000,10,0)”
//by mapping the importance of each attribute to weights based upon your business domain, you can easily find results which match your customer’s profile without the user having to initiate a search.
Hierarchical Recommendations• Example: Match User Attributes to Item Attribute Fields
Janes_Profile:{MostLikelyCategory:”healthcare//nursing//oncology”, 2ndMostLikelyCategory:”healthcare//nursing//transplant”, 3rdMostLikelyCategory:”educator//postsecondary//nursing”, …
}
/solr/select/?q=(category:((”healthcare.nursing.oncology”^40 OR ”healthcare.nursing”^20 OR “healthcare”^10))
OR (”healthcare.nursing.transplant”^20 OR ”healthcare.nursing”^10 OR “healthcare”^5))
OR (”educator.postsecondary.nursing”^10 OR ”educator.postsecondary”^5 OR “educator”) ))
Textual Similarity-based Recommendations
• Solr’s More Like This Request Handler / Search Handler are a good example of this.
• Essentially, “important keywords” are extracted from one or more documents and turned into a search.
• This results in secondary search results which demonstrate textual similarity to the original document(s)
• See http://wiki.apache.org/solr/MoreLikeThis for example usage
• Currently no distributed search support (but a patch is available)
Concept Based RecommendationsApproaches: 1) Create a Taxonomy/Dictionary to define your concepts and then either:
a) manually tag documents as they come in
or
b) create a classification system which automatically tags content as it comes in (supervised machine learning)
2) Use an unsupervised machine learning algorithm to cluster documents and dynamically discover concepts (no dictionary required).
//Very hard to scale… see Amazon Mechanical Turk if you must do this
//See Apache Mahout
//This is already built into Solr using Carrot2!
How Clustering Works
<searchComponent name="clustering" enable=“true“ class="solr.clustering.ClusteringComponent"> <lst name="engine"> <str name="name">default</str> <str name="carrot.algorithm">
org.carrot2.clustering.lingo.LingoClusteringAlgorithm</str> <str name="MultilingualClustering.defaultLanguage">ENGLISH</str> </lst></searchComponent> <requestHandler name="/clustering" enable=“true" class="solr.SearchHandler"> <lst name="defaults"> <str name="clustering.engine">default</str> <bool name="clustering.results">true</bool> <str name="fl">*,score</str> </lst> <arr name="last-components"> <str>clustering</str> </arr></requestHandler>
Setting Up Clustering in SolrConfig.xml
Clustering Search in Solr
• /solr/clustering/?q=content:nursing &rows=100 &carrot.title=titlefield &carrot.snippet=titlefield &LingoClusteringAlgorithm.desiredClusterCountBase=25 &group=false //clustering & grouping don’t currently play nicely
• Allows you to dynamically identify “concepts” and their prevalence within a user’s top search results
Search: Nursing
Search: .Net
Example Concept-based Recommendation
Clusters Identifier:Developer (22) Java Developer (13) Software (10) Senior Java Developer (9) Architect (6) Software Engineer (6) Web Developer (5) Search (3) Software Developer (3) Systems (3) Administrator (2) Hadoop Engineer (2) Java J2EE (2) Search Development (2) Software Architect (2) Solutions Architect (2)
Original Query: q=(solr or lucene)
// can be a user’s search, their job title, a list of skills, // or any other keyword rich data source
Stage 1: Identify Concepts
Facets Identified (occupation):Computer Software EngineersWeb Developers
...
Example Concept-based Recommendation
q=content:(“Developer”^22 or “Java Developer”^13 or “Software ”^10 or “Senior Java Developer”^9 or “Architect ”^6 or “Software Engineer”^6 or “Web Developer ”^5 or “Search”^3 or “Software Developer”^3 or “Systems”^3 or “Administrator”^2 or “Hadoop Engineer”^2 or “Java J2EE”^2 or “Search Development”^2 or “Software Architect”^2 or “Solutions Architect”^2) and occupation: (“Computer Software Engineers” or “Web Developers”)
// Your can also add the user’s location or the original keywords to the // recommendations search if it helps results quality for your use-case.
Stage 2: Run Recommendations Search
Example Concept-based Recommendation
Stage 3: Returning the Recommendations
…
Important Side-bar: Geography
Geography and Recommendations• Filtering or boosting results based upon geographical area or
distance can help greatly for certain use cases:– Jobs/Resumes, Tickets/Concerts, Restaurants
• For other use cases, location sensitivity is nearly worthless:– Books, Songs, Movies
/solr/select/?q=(Standard Recommendation Query) AND _val_:”(recip(geodist(location, 40.7142, 74.0064),1,1,0))”
// there are dozens of well-documented ways to search/filter/sort/boost // on geography in Solr.. This is just one example.
Behavior-based Recommendation Approaches(Collaborative Filtering)
The Lucene Inverted Index (user behavior example)
Term Documentsuser1 doc1, doc5user2 doc2user3 doc2user4 doc1, doc3,
doc4, doc5user5 doc1, doc4… …
Document “Users who bought this product” Field
doc1 user1, user4, user5
doc2 user2, user3
doc3 user4
doc4 user4, user5
doc5 user4, user1… …
What you SEND to Lucene/Solr:How the content is INDEXED into Lucene/Solr (conceptually):
Collaborative Filtering• Step 1: Find similar users who like the same documents
q=documentid: (“doc1” OR “doc4”)Document “Users who bought this
product “Fielddoc1 user1, user4, user5
doc2 user2, user3
doc3 user4
doc4 user4, user5
doc5 user4, user1… …
Top Scoring Results (Most Similar Users):1) user5 (2 shared likes) 2) user4 (2 shared likes)3) user 1 (1 shared like)
doc1
user1 user4 user5
user4 user5
doc4
• Step 2: Search for docs “liked” by those similar users
/solr/select/?q=userlikes: (“user5”^2 OR “user4”^2 OR “user1”^1)
Term Documentsuser1 doc1, doc5user2 doc2user3 doc2user4 doc1, doc3,
doc4, doc5user5 doc1, doc4… …
Collaborative Filtering
Top Recommended Documents:1) doc1 (matches user4, user5, user1)2) doc4 (matches user4, user5)3) doc5 (matches user4, user1)4) doc3 (matches user4)
//Doc 2 does not match//above example ignores idf calculations
Most Similar Users:1) user5 (2 shared likes)2) user4 (2 shared likes)3) user 1 (1 shared like)
Lot’s of Variations• Users –> Item(s)• User –> Item(s) –> Users• Item –> Users –> Item(s)• etc.
Note: Just because this example tags with “users” doesn’t mean you have to. You can map any entity to any other related entity and achieve a similar result.
User 1 User 2 User 3 User 4 …Item 1 X X X …Item 2 X X …Item 3 X X …Item 4 X …… … … … … …
Comparison with Mahout• Recommendations are much easier for us to perform in Solr:
– Data is already present and up-to-date– Doesn’t require writing significant code to make changes (just changing queries)– Recommendations are real-time as opposed to asynchronously processed off-line.– Allows easy utilization of any content and available functions to boost results
• Our initial tests show our collaborative filtering approach in Solr significantly outperforms our Mahout tests in terms of results quality– Note: We believe that some portion of the quality issues we have with the Mahout implementation
have to do with staleness of data due to the frequency with which our data is updated.
• Our general take away:– We believe that Mahout might be able to return better matches than Solr with a lot of custom work,
but it does not perform better for us out of the box.
• Because we already scale…– Since we already have all of data indexed in Solr (tens to hundreds of millions of documents), there’s
no need for us to rebuild a sparse matrix in Hadoop (your needs may be different).
Hybrid Recommendation Approaches
Hybrid Approaches• Not much to say here, I think you get the point.
• /solr/select/?q=category:(”healthcare.nursing.oncology”^10 ”healthcare.nursing”^5 OR “healthcare”) OR title:”Nurse Educator”^15 AND _val_:”map(salary,40000,60000,10,0)”^5 AND _val_:”(recip(geodist(location, 40.7142, 74.0064),1,1,0))”)
• Combining multiple approaches generally yields better overall results if done intelligently. Experimentation is key here.
Important Considerations & Advanced Capabilities @ CareerBuilder
Important Considerations @ CareerBuilder
• Payload Scoring• Measuring Results Quality• Understanding our Users
Custom Scoring with Payloads• In addition to boosting search terms and fields, content within the same field can also be boosted
differently using Payloads (requires a custom scoring implementation):
• Content Field:design [1] / engineer [1] / really [ ] / great [ ] / job [ ] / ten[3] / years[3] / experience[3] / careerbuilder [2] / design [2], …
Payload Bucket Mappings:jobtitle: bucket=[1] boost=10; company: bucket=[2] boost=4;
jobdescription: bucket=[] weight=1; experience: bucket=[3] weight=1.5
We can pass in a parameter to solr at query time specifying the boost to apply to each bucket i.e. …&bucketWeights=1:10;2:4;3:1.5;default:1;
• This allows us to map many relevancy buckets to search terms at index time and adjust the weighting at query time without having to search across hundreds of fields.
• By making all scoring parameters overridable at query time, we are able to do A / B testing to consistently improve our relevancy model
Measuring Results Quality• A/B Testing is key to understanding our search results quality.
• Users are randomly divided between equal groups
• Each group experiences a different algorithm for the duration of the test
• We can measure “performance” of the algorithm based upon changes in user behavior:– For us, more job applications = more relevant results– For other companies, that might translate into products purchased, additional friends
requested, or non-search pages viewed
• We use this to test both keyword search results and also recommendations quality
Understanding our Users (given limited information)
Understanding Our Users• Machine learning algorithms can help us understand what
matters most to different groups of users.
Example: Willingness to relocate for a job (miles per percentile)
1% 5% 10% 20% 25% 30% 40% 50% 60% 70% 75% 80% 90% 95%0
500
1,000
1,500
2,000
2,500
Title Examiners, Abstractors, and Searchers
Software Developers, Systems Software
Food Preparation Workers
Key Takeaways
• Recommendations can be as valuable or more than keyword search.
• If your data fits in Solr then you have everything you need to build an industry-leading recommendation system
• Even a single keyword can be enough to begin making meaningful recommendations. Build up intelligently from there.
Contact Info
And yes, we are hiring – come chat with me if you are interested.
Trey [email protected]://www.careerbuilder.com@treygrainger