AI to create professional
opportunities
Liang Zhang
Director of Engineering, AI
AI Tech Meetup 5/9/19
We are all seeking new professional opportunities
To Advance our Careers
Let people know you’re open
LinkedIn operates the largest professional network on the Internet
630M+ members
50K+ skills
25K+
titles
200+
countries
30M+ orgs
148 industries
certificates,
degrees, and
more ...states, cities,
postal codes, ...
roles, occupations
speciality
tools, products,
technologies, ...
Tell your
Story
AI is like oxygen at LinkedIn
People You May Know Feed
Jobs Learning
Recruiter Search Sales
Our Scale
25 B
ML A/B experiments
per week
data processed offline
per day
2002.15 PB
data processed
nearline per day
2 PB
graph edges with 1B
nodes
53 B
parameters in ML
models
Our Technology
Personalization at Scale
Personalization models creates significant value for LinkedIn
10
Basic model: Feature-based wide-and-deep
▪ Good for modeling general user behavior
▪ Not effective to capture the idiosyncratic
behavior of individual users and individual items
Basic model with real-time statistics
as features:
▪ Examples of real-time statistics:
– #clicks on item i
– #times user u clicked items of
topic k
▪ Good for statistics with sufficient
sample sizes
▪ Not effective when the sample
sizes of the statistics are small =>
high variance, often required to
capture segmented popularity
Linear Neural Net Trees
A class of flexible model that we find useful for personalization
Global Model
(Macro)
Per-User Model
(Micro)
Per-Item Model
(Micro)
Learns general user
behavior using a wide-
and-deep model with real-
time statistics.
Example: Users like to read
about articles on topics
related to their skills
Learns the behavior of an
individual user
One model per user
Example: Deepak likes to
read about articles on
startups, IoT and Cloud
Deepak’s own model has
100s of parameters to
capture his particular
behavior
Learns the behavior of an
individual item
One model per item
Example: Article on Kai-Fu
Lee’s book is liked by people
in Silicon Valley with
Marketing skills
Kai-Fu’s article has 100s of
parameters
Generalized Additive Mixed Effect Model
Reinforcement Learning with Immediate Rewards
Perform reinforcement learning with immediate rewards via Thompson Sampling to collect the most
useful data for the micro models
Evaluate models offline using historical replay on randomized data (with inverse-sampling-rate
weighting) to mitigate bias in data collection
Current Work
▪ Learning nonparametric priors for per-entity models via Neural embedding models
▪ Full reinforcement Learning to directly optimize for long-term rewards like DAU, confirmed hires, etc
12
Global
Model
Per-User
Model
Per-Item
Model
Estimate the posterior distribution of model parameters and draw a score
from the distribution
Search @ LinkedIn
▪ Traditional Search Engine (Google, Bing…): MaxDocSim <Query, Doc>
▪ LinkedIn Search Engines (Personalized): MaxDocSim <Query, Doc> | Person (social network, profile,
interactions…)
▪ Approach: Traditional IR + NLP via Deep Neural Network + Personalization
Search Personalization
Video @ LinkedIn
Feed Search Learning
Content Quality
Computer Vision + Speech Recognition + NLP
15
Pro-ML: Productive Machine Learning
▪ Make the end-to-end process of running and iterating on large ML workflows easy, robust and almost automated
Model
Deployment
Model
Maintenance
Feature
Engineering
Target
Definition
Model
Creation
#Experiments per Eng
Business
Impact
Core Team
16
SNVSunnyvale
SFSan Francisco
DUB Dublin
BLR Bangalore
NYC New York City
500+
Eng.
We are hiring!

AI in linkedin

  • 1.
    AI to createprofessional opportunities Liang Zhang Director of Engineering, AI AI Tech Meetup 5/9/19
  • 2.
    We are allseeking new professional opportunities To Advance our Careers Let people know you’re open
  • 3.
    LinkedIn operates thelargest professional network on the Internet 630M+ members 50K+ skills 25K+ titles 200+ countries 30M+ orgs 148 industries certificates, degrees, and more ...states, cities, postal codes, ... roles, occupations speciality tools, products, technologies, ... Tell your Story
  • 4.
    AI is likeoxygen at LinkedIn
  • 5.
    People You MayKnow Feed
  • 6.
  • 7.
  • 8.
    Our Scale 25 B MLA/B experiments per week data processed offline per day 2002.15 PB data processed nearline per day 2 PB graph edges with 1B nodes 53 B parameters in ML models
  • 9.
  • 10.
    Personalization at Scale Personalizationmodels creates significant value for LinkedIn 10 Basic model: Feature-based wide-and-deep ▪ Good for modeling general user behavior ▪ Not effective to capture the idiosyncratic behavior of individual users and individual items Basic model with real-time statistics as features: ▪ Examples of real-time statistics: – #clicks on item i – #times user u clicked items of topic k ▪ Good for statistics with sufficient sample sizes ▪ Not effective when the sample sizes of the statistics are small => high variance, often required to capture segmented popularity Linear Neural Net Trees
  • 11.
    A class offlexible model that we find useful for personalization Global Model (Macro) Per-User Model (Micro) Per-Item Model (Micro) Learns general user behavior using a wide- and-deep model with real- time statistics. Example: Users like to read about articles on topics related to their skills Learns the behavior of an individual user One model per user Example: Deepak likes to read about articles on startups, IoT and Cloud Deepak’s own model has 100s of parameters to capture his particular behavior Learns the behavior of an individual item One model per item Example: Article on Kai-Fu Lee’s book is liked by people in Silicon Valley with Marketing skills Kai-Fu’s article has 100s of parameters Generalized Additive Mixed Effect Model
  • 12.
    Reinforcement Learning withImmediate Rewards Perform reinforcement learning with immediate rewards via Thompson Sampling to collect the most useful data for the micro models Evaluate models offline using historical replay on randomized data (with inverse-sampling-rate weighting) to mitigate bias in data collection Current Work ▪ Learning nonparametric priors for per-entity models via Neural embedding models ▪ Full reinforcement Learning to directly optimize for long-term rewards like DAU, confirmed hires, etc 12 Global Model Per-User Model Per-Item Model Estimate the posterior distribution of model parameters and draw a score from the distribution
  • 13.
    Search @ LinkedIn ▪Traditional Search Engine (Google, Bing…): MaxDocSim <Query, Doc> ▪ LinkedIn Search Engines (Personalized): MaxDocSim <Query, Doc> | Person (social network, profile, interactions…) ▪ Approach: Traditional IR + NLP via Deep Neural Network + Personalization Search Personalization
  • 14.
    Video @ LinkedIn FeedSearch Learning Content Quality Computer Vision + Speech Recognition + NLP
  • 15.
    15 Pro-ML: Productive MachineLearning ▪ Make the end-to-end process of running and iterating on large ML workflows easy, robust and almost automated Model Deployment Model Maintenance Feature Engineering Target Definition Model Creation #Experiments per Eng Business Impact
  • 16.
    Core Team 16 SNVSunnyvale SFSan Francisco DUBDublin BLR Bangalore NYC New York City 500+ Eng.
  • 17.

Editor's Notes

  • #3 That’s why we are all here. Some of us want to present our research and findings and build reputation, some of us want to enhance our skills and stay informed to become better at our jobs, some of us are here seeking new job opportunities, some of us just want to build and nurture our network for the future, and so on.
  • #4 It is a platform for every professional to tell their story. Who they are, where they work, skills, etc. Once you have done that, the platform works hard all the time and helps you connect to opportunity.
  • #6 Upated PYMK, updating feed in one moment
  • #7 Learning okay, Jobs updated
  • #16 The focus is on iterating the ML models fast.