CS-RAD: Conditional Member Status Refinement and Ability Discovery for Social Network Applications

Yiming Ma�Knowledge Graph, LinkedIn

​

KDD 2022

Agenda

Definition of Status & Ability

Refinement & Gap Discovery Problems

Algorithms

Experiments

1

2

3

4

Acknowledgements

I would like to express my gratitude and appreciations to LinkedIn

Data Organization management team, especially Qi He, Romer

Rosales-Delmoral, and William Wei Kang in supporting this R&D

effort; to LinkedIn product team, especially Anurag Aggarwal in

providing LinkedIn product perspectives to this research; to my colleagues who have provided valuable early feedback to this research,especially Yiping Yuan, Dash Shi, Yi Zhen, and Yiyang Hao

​

Member Status & Abilities in a Social Network

• Status represents a member’s social value in the network

​

• Abilities represent the level of credibility and authority for a member to hold certain status

Status matches well with ability

Status:

​

Professional Title: CEO at LinkedIn

Ability:

Abilities More Than Status

Status More Than Abilities

Status Refinement Problem

• Can we refine a Source title to one of the possible target titles based on ability vector E ?

​

RefinementScore (st → tt) = Pr(tt|E, st)

​

​

Title Refinement Hierarchy (H)

Edge Modeling

• Build models for all edges in the refinement Hierarchy H
• Intuition: we provide context to refine a title based on its parent title

​

• Edge Key:
• Parent Node: Source Title (st) or Less specific title
• Child Node: Target title (tt) or More specific title
• Model Parameters:
• Set of Skills (E) with Weights
• With Edge Model, we can infer:
• Pr(tt|E, st)

Apply Edge models (inference) for Title Refinement

• A member with current title st in H
• For all edge models with the children tt of st:
• Choose tt with highest Pr(tt|E, st):
• Further filter by a predefined threshold (e.g., 95%) to ensure good match:

​

Ability Discovery Problem

• Skills are one of the most important types of Ability at LinkedIn

​

• Given a target title 𝑡𝑡 and current skill set 𝐸
1. Assess the goodness of the skill set 𝐸
2. identify the Skill gap

Apply Edge models (inference) for Skill Discovery

• A member with current title tt in H
• For all edge models with the parents st of tt:
• If goodness measure Pr(tt|E, st) < threshold
• Add most relevant skills E’ to increase Pr(tt|E+E’, st) > threshold
• Example: tt: Sales Manager

Edge Modeling: Mixture of Child Titles

• Source title st (parent) can be expressed by mixture of Target title tt and other titles: Pr(st|tt) + Pr(other|tt) = 1.0

​

Edge Modeling: Mixture of skills

• A parent skill is a mixture of the child skills

​

​

Edge Modeling: Refinement Probabilistic Model

• Assume conditional independence of skills

​

​

Log of Mixture Ratio

Log Odds of a skill e_i

Edge Modeling: Refinement Probabilistic Model

• Add missing positive skills will increase Pr(tt|E, st)

​

​

Experiment Setup

Model Inspection

CS-RAD

TS-Dist: Title & Skill Distribution

Model Inspection

LDA: Skill topics for Manager title

Refinement Behavior

Gap Discovery Behavior

Gap Discovery Behavior

Future work

• Move CS-RAD to production and help millions of members around the world improve their social status and abilities

​

• Extending our status and ability beyond Titles and Skills

Thank you!