ML Platform Q1 Meetup: An introduction to LinkedIn's Ranking and Federation Libraries
- 1. Quasar and ReMix An introduction to LinkedIn's Ranking and Federation libraries Andris Birkmanis & Lance Wall 1
- 2. Relevance: Verticals & Infrastructure 2 Relevance Isolated ML models Integrated ML models Relevance Infra Relevance Verticals Deployed ML services ML algos Scoring and Ranking Tools Relevance service platform Quasar ReMix
- 3. Quasar Quick Scoring and Ranking Our mission is making efficient feature transformation, scoring, and ranking simple. 3
- 4. Scoring • Scoring • Scorables • Features • Feature Transformation 4
- 5. Ranking • Sorting • TopK • Filtering • Group by • Distinct • Union • Custom 5
- 6. Relevance Models: DAGs of Computations Filter BY interest-match > 0.5 Filter BY skill-match > 0.7 TOP 50 BY content-match All Documents member interest category interest- match- score news feed skill content skill- match- score content- match- score 10,000 500 500 50
- 7. ML and Training • Tracking training dependencies between ML models • Integrating with training engines via Training API • Automatic type conversion for features and model parameters • Reuse of feature transformations between training and prediction 7
- 8. Quasar Components • Domain Specific Language (DSL) ▪ Oriented towards scoring and ranking concepts ▪ Supports various machine learning models ▪ Supports various ranking operators ▪ Supports pluggable feature transformers ▪ Supports arithmetical and logical expressions • Library ▪ Includes out-of-box feature transformers tuned for performance (dense/sparse vectors, bags of words, etc.) ▪ Extensible with custom transformers and ranking operators • Execution engine ▪ Supports multiple evaluation strategies for different objectives (lazy/eager/batching/etc.) ▪ Debuggability, logging, and other cross-cutting concerns ▪ API for scoring, ranking, read/write access to features, training
- 9. • LinkedIn Relevance Products ▪ Feed, Recommendation, Search • Adoption ▪ 1000+ Quasar models Project Status
- 10. Future directions • Better training support for external models (XGBoost, Tensorflow) • Making feature transformers and operators more reusable • Better type information • Standardized storage formats for features and model parameters • See the upcoming LinkedIn engineering blog for technical details 10
- 11. Lance Wall ReMix
- 12. Example relevance workflows at LinkedIn Member ID Fetch Member Profile Fetch Member Profile Compute Job Recommendations Compute People Recommendations Format Results Member ID Format Results
- 13. Motivation • Multitenant relevance workflow services with tens of engineers on multiple teams contributing • Each relevance workflow service has different APIs and conventions • Lack of abstraction of system-level concerns from application logic • Diminished productivity, operability, and leverage
- 14. ReMix’s Mission Provide an easy to use platform for building relevance services with a focus on optimizing leverage and automating common operability concerns.
- 15. Design Goals • Consolidation of various relevance service stacks • Ease of support • Ease of development • Ease of operation
- 16. Features of ReMix • Leverages ParSeq for easy asynchronous I/O • Exposes declarative API for composing workflows • Provides automated monitoring instrumentation and tooling • Provides robust, extensible solutions for common workflow functionality • Provides isolation and robustness to downstream instability
- 17. How does ReMix work? Operator is assembled into Workflow is submitted to WorkflowEngine
- 18. Operator • Modular functional component of a Workflow • ReMix provides Operators for common functionality • ReMix provides decorative interfaces for common optimizations • ReMix provides generic support for asynchronous execution
- 19. Example relevance workflows at LinkedIn Member ID Fetch Member Profile Fetch Member Profile Compute Job Recommendations Compute People Recommendations Format Results Member ID Format Results
- 20. Workflow • Declaration of deferred execution • Easy to understand declarative language • Leverages ParSeq and exposes a simpler API • Abstraction of execution behavior & optimizations • Independent of environment or service (i.e. portable)
- 21. Example relevance workflows at LinkedIn Member ID Fetch Member Profile Fetch Member Profile Compute Job Recommendations Compute People Recommendations Format Results Member ID Format Results
- 22. WorkflowEngine • Executor of Workflows • Translates Workflows to ParSeq Tasks • Provides special considerations for async/RPC operations • Provides common operability functionality
- 23. Project Status & Planned Work • ReMix adopters include job recommendations and blended search • Working on integration with Quasar ▪ Complete solution for model serving from offline to online • ReMix Cloud ▪ Simple toolkit/UI for creating a Workflow and deploying it to production ▪ Hosts Workflows in a managed service, with little to no operational cost to Workflow developers ▪ Increased leverage due to reuse of common components in multitenant platform
- 24. Thanks!
- 25. Backup Slides 25
- 26. Quasar Model 26 MODELID "feed_quasar"; DOCPARAM com.linkedin.feed.FeedItem feedItem; REQUEST PARAM Profile member; REQUEST FEATURE VECTOR interests = GetInterests(member); DOCUMENT FEATURE VECTOR categories = GetCategories(feedItem); DOCUMENT FEATURE LONG publishedTime = GetPublishedTime(feedItem); MODEL PARAM timeBuckets = { "1hr" : 60, "3hr" : 180 }; DOCUMENT FEATURE VECTOR normalizedTime = Bucketize(diffTime, timeBuckets); DOCUMENT FEATURE VECTOR interestMatch = Similarity(interests, categories); MODEL PARAM MAP<STRING, OBJECT> modelWeights = { “normalizedTime”: { “1hr”: 0.234, “3hr”: 0.456, “Other”:0.21 }, “interestMatch”: 0.823 }; DOCUMENT FEATURE FLOAT score = LinearScore(modelWeights, “sigmoid”); DOCUMENT FEATURE BOOLEAN aboveThreshold = score > 0.5 filteredFeed = FILTER DOCUMENTS BY aboveThreshold; rerankedFeed = ORDER filteredFeed BY score WITH DESC; RETURN rerankedFeed;
- 27. Candidate list of documents Filter Documents getInterest s getCateg ories getPublish edTime getSimilari ty Bucketize LinearSco re getCateg ories getPublish edTime getSimilari ty Bucketize LinearSco re getCateg ories getPublish edTime getSimilari ty Bucketize LinearSco re getCateg ories getPublish edTime getSimilari ty Bucketize LinearSco re 1 3 4Request 1 3 4 3 1 4 Order Documents Pass 1 2 Pass 2 Decision Tree LinearSc ore Decision Tree LinearSc ore getVie wTimes Bucke tize Decision Tree LinearSc ore The multipass ensemble model at runtime
- 28. Vector Math and Expression Support • Vector as first class citizen in DSL • State-of-art Java Vector implementation ▪ Compact and efficient data structure ▪ Efficient Vector math computation C++ Java Networ k 1.0 1.0 3.0 1.0 Linux Member/Job Similarity Score = member.skill Hadoop Scala Gradle 2.0 1.0 2.0 job.required_skill dot product