Aiinpractice2017deepaklongversion
Download as PPTX, PDF4 likes1,390 views
The document discusses how AI is used at scale to create professional opportunities. It provides an overview of how AI powers the user and customer experience on LinkedIn through search, recommendations, staying informed, and getting hired. It describes how AI uses profile and network data to improve recommendations through understanding member characteristics and connections. The document also discusses how LinkedIn's recommendation system works, including using a generalized additive mixed-effect model called GLMix for large-scale regression to provide personalized job recommendations.
Aiinpractice2017deepaklongversion
- 1. Deepak Agarwal AI in practice, AAAI, San Francisco, USA Feb 6th, 2017 AI that creates professional opportunities at scale
- 2. To Advance our Careers We are seeking new professional opportunities all the time
- 4. Emily likes AI and deep learning. Got familiar with research on Dropout through an article shared by one of her connection. Followed through, implemented and got great recognition at work.
- 5. Impact of creating professional opportunities at scale • Closing the skills gap by leveraging supply on a global basis, reducing income disparity, training the workforce to prepare for the future. JOBS QUEUE in INDIA, over supply
- 6. Almost all user experience powered by AI SEARCH,CONNECT, NURTURE Search and connect, via recommendations, keep in touch STAY INFORMED through professional news and knowledge GET HIRED and build your career
- 7. All customer experience too HIRE Searching, researching, nurturing qualified candidates MARKET Targeting right audience via native ads, display ads SELL Searching and researching decision makers, following up on leads
- 8. How is AI used to drive value ? • Two broad classes of problems •Recommendation : ◆Recommend items to a users to optimize one of more objectives of interest (e.g., connections, job applies, engagement/revenue) •Search: ◆User asks a question through query, need to provide answer by searching through a large set of possibilities
- 10. Connecting long-term objectives to proxies that can be optimized by machines/algorithms Long-term objectives (return visits to site, connections, quality job applies,,..) Short-term proxies (CTR, connection prob, apply prob, …) Large scale optimization via ML.. Experiment Learn Innovate
- 11. Recommendation Problem: Not merely supervised learning User i with user features xi (profile info, browse history, geo-location, search history, …) visits item j with item features xj (keywords, content categories, author, ...) Algorithm selects (i, j) : response yij Interaction (Click, share, like, apply, connect,..)/no-interaction Which item should we select? • The one with highest predicted utility • The one most useful for improving the utility prediction model Exploit Explore
- 12. Data: Profile and Network Key to better AI
- 13. User Characteristics Profile Information Title, skills, work history, education, endorsements, presentations, follows,.. Behavioral Activities, search,.. Professional profile of record
- 14. AI used to standardize member profile data text ID ● language detection ● spell correction ● segmentation ● entity recognition ● entity resolutionprofile IDs in taxonomy text mining machine learning
- 15. An AI-based holistic solution for taxonomy development human-curation to create and structure taxonomies slow, costly, not scalable deep neural network for entity embedding Auto-generate ● new entities ● synonyms ● entity hierarchical structure etc. fast, cheap, scalable
- 16. Proactively helping members with profile completion Estimate the rewards of adding a new skill through peer analysis Infer the skill reputation of LinkedIn member through user profiles (static), user behaviors (dynamic) and user connections (network) Auto-generate profile summary through knowledge base
- 17. Better knowledge base powers better AI Members (profile), Jobs, Companies etc. collectively form the LinkedIn professional knowledge base direct knowledge base approach based on logical inference rules direct features other training data like network, user behaviors graph embedding better representation Search Job recommendation Profile edit Ads targeting Feed Course learning Insights Endorsement …...
- 18. Network: Connections are fundamental to drive value • A well connected member is much more likely to be able to use the LinkedIn eco-system for her professional development and growth 18
- 20. Adding nodes: Viral Loops through recommendations C Landing Page ✔ ✔ Contacts Upload Abc abc@myschool.edu XYZ, lXYZ@myschool.edu ✔
- 21. Adding nodes: Viral loops from Guest Invitation • Members inviting others to join the network is crucial for growth of the network • Anderson et. al. (WWW 2015) • LinkedIn’s cascade trees are very viral 21 The growth of trees is almost linear in time
- 22. Item Features Articles author, sharer, keywords, named entities, topics, category, likes, comments, latent representation, etc. Job Post company, title,skills, keywords, geo, … .......
- 23. User Intent • Why are you here ? •Hire, get hired, stay informed, grow network, nurture connections, sell, market,.. •Explicit (e.g., visiting jobs homepage, search query), •Implicit like job seeker (needs to be inferred, e.g., based on activities)
- 24. We know about users, their intent, about items How to Recommend items to users algorithmically ? Framework, Infrastructure and Tools
- 25. Under the Hood of a Typical Recommender System at LinkedIn 25
- 26. Example Application: Job Recommendation
- 27. Objective: Job Applications Predict the probability that a user i would apply for a job j given … •User features •Profile: Industry, skills, job positions, companies, education •Network: Connection patterns •Item (i.e., job) features •Source: Poster, company, location •Content: Keywords, title, skills •Data about users’ past interactions with different types of items •Items: Jobs, articles, updates, courses, comments •Interactions: Apply, click, like, share, connect, follow
- 28. Components Front End Service Ranking Service Item Index User Feature Stores User DB Item DB Offline Data Pipelines Item Feature Pipelines User Feature Pipelines Data Stream Processing User Activity Data Streams Live Index Updater ETL ETL Online Offline Model Training Pipelines Offline Index Builder User Photon-ML Apache Hadoop, Pig, Scalding, SparkSearch Index Experimentation Platform A/B testing Ranking Library
- 29. Model Training Raw User Features Raw Item Features DAG of Transformers DAG of Transformers DAG of Transformers Feature Vector of User i xi Matching Feature Vector mij Feature Vector of Item j zj (trees, similarities) Parameter vector for each user i Parameter vector for each item j p(i applies for j) = f( xi, zj, mij | θ, αi, βj ) Feature Processing Parameter Learning Global parameter vector
- 30. Model Deployment User Feature Stores Live Index UpdaterItem Index Parameter vector for each user i Parameter vector for each item j p(i applies for j) = f( xi, zj, mij | θ, αi, βj ) Global parameter vector
- 31. Online Ranking User Feature Stores User DB User Feature Pipelines Data Stream Processing User Activity Data Stream Ranking Service Item Index Offline Data Pipelines ETL ETL Online Offline Model Training Pipelines Offline Index Builder Front End Service User User Features & Parameters Item DB Live Index Updater Item Feature Pipelines
- 32. Online A/B Experiments Experiment setting - Select a user segment - Allocate traffic to different models Result reporting - Report experimental results - Impact on a large number of metrics
- 33. How we do large scale Regression ? GAME: Generalized Additive Mixed-Effect Model
- 34. GAME as a framework, optimizer and library
- 35. GAME as a Framework • Unifies and mixes different models into a principled additive model. • Predicting the response of user i on item j: • : An effect (model) • : Expectation of response • : Link function
- 36. GAME as a Library • Basic models implemented as building block . • Matrix factorization • Generalized Linear Model • … • New models can be directly composed by mixing existing building blocks.
- 37. Ex: GLM + MF • Predicting the response of user i on item j: • : GLM • : MF • : Expectation of response • : Link function
- 38. GLMix¹: Fine-Grained GAME with Linear Components 1: GLMix: Generalized Linear Mixed Models For Large-Scale Response Prediction X. Zhang et al., KDD2016
- 39. • Jobs homepage • Ramped to serve 100% traffic (400 million LinkedIn members) • significant lift in job application rate 41 GLMix @
- 40. • Jobs homepage • Ramped to serve 100% traffic (400 million LinkedIn members) • significant lift in job application rate • Advertising, PYMK, Feed, Recruiter,… • significant lifts in performance 42 GLMix @
- 41. Generalized Linear Model (GLM) • Predicting the response of user i on item j: • : Feature vector (profile, items, intent, cross-prod, GBDT) • : Coefficient vector • : Expectation of response • : Link function
- 42. ●Alice and Annie are about the same age, similar majors in college… (similar member features ) ●Alice likes to take more risks with start-ups ●Annie likes more stable career just like her parents ●GLM may return similar set of jobs to both GLM for Job Recommendation 44
- 43. ●Alice and Annie are about the same age, similar majors in college… (similar member features ) ●Alice likes to take more risks with start-ups ●Annie likes more stable career just like her parents ●GLM may return similar set of jobs to both ●Need more fine-grained modeling at different granularity to better personalize the model! GLM for Job Recommendation 45
- 44. GLMix: Generalized Linear Mixed Model • Predicting the response of user i on item j: • Model coefficients with different granularities: • Per-user random effect coefficients • Per-item random effect coefficients • GLMix = GLM + per-user model + per-item model
- 45. GLMix for Job Recommendation ● Global fixed effect model ○ Similarity between member profile and jobs profile, e.g. do the member skills and job skills look similar?
- 46. GLMix for Job Recommendation ● Global fixed effect model ○ Similarity between member profile and jobs profile, e.g. do the member skills and job skills look similar? ● Per-member random effect model ○ E.g. If a member has applied to a job with title = “software engineer”, we will boost “software engineer” jobs more in her results.
- 47. GLMix for Job Recommendation ● Global fixed effect model ○ Similarity between member profile and jobs profile, e.g. do the member skills and job skills look similar? ● Per-member random effect model ○ E.g. If a member has applied to a job with title = “software engineer”, we will boost “software engineer” jobs more in her results. ● Per-job random effect model ○ E.g. If a job gets an apply with a member titled “software engineer”, we will boost this job more for members with this title.
- 48. Alice and Annie’s problem revisited ● Per-user random effect coefficients for Alice: ● Per-user random effect coefficients for Annie: ● Alice and Annie now may have different job recommendations given their per-user coefficients.
- 49. Fine-Grained Modeling with GAME
- 50. Takeaways • GAME unifies and mixes different models into a principled additive model. • MF + GLM = RLFM/Matchbox • GLM + DNN = Wide & Deep Learning • ... 52
- 51. Takeaways • GAME unifies and mixes different models into a principled additive model. • MF + GLM = RLFM • GLM + DNN = Wide & Deep Learning • ... • GLMix is GAME with linear component that captures signal from different granularity • GLMix = GLM + Per-member model + per-item model + … 53
- 52. Takeaways • GAME unifies and mixes different models into a principled additive model. • MF + GLM = RLFM • GLM + DNN = Wide & Deep Learning • ... • GLMix is GAME with linear component that captures signal from different granularity • GLMix = GLM + Per-member model + per-item model + … • GAME is part of an open-source library • Search for Photon-ML • https://github.com/linkedin/photon-ml 54
- 54. Some Challenges and Learnings
- 55. 1. Cost of a Bad Recommendation • Making ML robust when a few bad recommendations can hurt product brand • Maximize precision without hurting performance metrics significantly • Collect negative feedback from users, crowd; incorporate within algorithms • Create better product focus, filter unnecessary content from inventory • E.g., unprofessional content on Feed • Better insights/explanations associated with recommendations help build trust
- 56. 2. Data Tracking • Proper data tracking and monitoring is not always easy! • Data literacy and understanding across organization (front-end, UI, SRE) • Proper tooling, continuous monitoring very important to scale the process • Our philosophy: Loose coupling between FE and BE teams! • FE (client) emits limited events along with trackingid • BE emits more details and joins against trackingid • Tracking events can have significant impact • View-port tracking (what the user actually saw) for more informative negatives
- 57. 3. Content Inventory • High quality and comprehensive content inventory as important as recommendation algorithms •Examples: Jobs, Feed •Supply and demand analysis, gap analysis, proactively producing more high quality content for creating appropriate inventory
- 58. 4. A/B Testing with Network Interference • Random treatment assignments (spillover effects, need to adjust) • Treatment recommendations affect control group as well • A like/share in treatment may create a new item when ranking in control 60