Overview
The article discusses the development of a heterogeneous social network recommendation system at LinkedIn, focusing on the 'People You May Know' feature. It highlights the challenges and strategies involved in ranking diverse recommendations based on different edge types and entities.
What You'll Learn
1
How to implement a two-phased ranking process for heterogeneous recommendations
2
Why calibration of Edge-FPR model scores is crucial for consistent ranking
3
How to leverage counterfactual experiments to estimate edge importance
Prerequisites & Requirements
- Understanding of recommendation systems and machine learning concepts
- Familiarity with XGBoost and logistic regression(optional)
Key Questions Answered
What are the different edge types in LinkedIn's recommendation system?
LinkedIn's recommendation system includes three edge types: CONNECTION edges, which are bidirectional and allow members to connect; FOLLOW edges, which are unidirectional and allow members to follow hashtags, companies, or creators; and SUBSCRIBE edges, which enable members to subscribe to groups, newsletters, or events. Each edge type serves distinct purposes in building a heterogeneous social network.
How does the SPR algorithm rank heterogeneous cohorts?
The SPR algorithm ranks heterogeneous cohorts through a two-phase process. First, it ranks entities of one type among themselves using Edge-FPR models. Then, it ranks these heterogeneous cohorts against each other to help members select the next edge to grow their network. This process ensures a comprehensive recommendation system that considers various types of entities.
What challenges are faced in developing the recommendation system?
Developing the recommendation system involves three main challenges: calibrating scores from different Edge-FPR models for consistent ranking, addressing disproportionate behavior in member interactions across edge types, and managing the coupling between SPR and Edge-FPR models to maintain development agility. Each challenge requires careful consideration to optimize the recommendation process.
Key Statistics & Figures
Increase in engaged members
Significant increase
Measured through A/B tests conducted after implementing the SPR system.
Downstream interactions
More members creating edges and having conversations
This was a direct result of the improved recommendation system.
Technologies & Tools
Machine Learning
Xgboost
Used in the SPR algorithm to predict the probability of downstream interactions.
Key Actionable Insights
1Implement a two-phased ranking system to enhance recommendation accuracy.This approach allows for a more nuanced understanding of user preferences by first ranking similar entities and then comparing different types of recommendations, leading to better user engagement.
2Regularly calibrate Edge-FPR model scores to ensure comparability.Calibration is essential for maintaining a consistent ranking across diverse edge types, which can significantly impact user experience and satisfaction.
3Utilize counterfactual experiments to assess the true value of edge types.By temporarily removing certain edges and observing user engagement, you can derive importance factors that enhance the recommendation system's effectiveness.
Common Pitfalls
1
Failing to calibrate scores from different Edge-FPR models can lead to inconsistent recommendations.
Without proper calibration, the ranking of heterogeneous entities may become skewed, resulting in poor user experience and engagement.
2
Overrepresenting follow edges due to member behavior can distort the recommendation balance.
If the training data is biased towards follow edges, it may lead to suboptimal recommendations for connection edges, which could be more valuable for users.
Related Concepts
Recommendation Systems
Machine Learning Algorithms
Social Network Analysis
User Engagement Strategies