This document summarizes LinkedIn's approach to personalizing users' news feeds using machine learning. It discusses how the feed helps users stay informed, connected, and establish their brand. It then describes LinkedIn's feed recommendation system and the challenges of recommending heterogeneous content to different user segments. The rest of the document outlines LinkedIn's feed relevance and mixer systems, the features and models used, and techniques for addressing issues like joint feed effects and ensuring diversity.

1. Personalizing LinkedIn Feed
Guy Lebanon
September 24, 2015
Collaborators: Deepak Agarwal, Kevin Chang, Bee-Chung Chen,
Boyi Chen, Qi He, Zhenhao Hua, Yiming Ma, Mikhail Obukhov,
Pannagadatta Shivaswamy, Liang Tang, Hsiao-Ping Tseng, Jaewon
Yang, Lin Yang, and Liang Zhang

2. LinkedIn Feed: Value and Impact
Main gateway to LinkedIn’s website and mobile app.
Help users stay informed by showing news and other posts
related to their career
Help users stay connected to their professional network by
showing updates from connections
Help users establish their brand/reputation via sharing and
posting.
In addition, help user find jobs (JYMBII), grow network
(PYMK), and generate ad revenue.
Small tweaks make huge impact on hundreds of millions of
LinkedIn members.

4. LinkedIn Feed: A Machine Learning Problem
Recommendation system with a few twists:
Heterogenous inventory: updates from my connections, jobs
recommendations, news recommendations, ads, people you
may know, etc.
Explicit and Implicit signals: clicks, like, share, comment, VPT
Different value for different member segment: job seeker,
novice, consumer
High quality profile
Social graph data
Spam and unprofessional content
Online scoring of a very large number of candidate updates

5. Feed Relevance System
multiple homogenous sources =⇒
each rank top-k =⇒
combine top-k lists into a single ranked list =⇒
reranking stage

6. Feed Mixer System
Feed mixer combines top-k ranked lists
Online component responds to web server request
Offline component prepares training data, trains model, process
features

7. Activities
Most feed activities are formulated as (actor, verb, object) triplet,
whose type (actor type, verb type, object type) is the activity type.

8. Model
Ranking is based on logistic regression model capturing
interactions (clicks, likes, etc.). If yit is a binary variable
capturing whether user i interacted with update t, then
P(yit = 1 | user, update) =

1 + exp


j
βj [Xit]j




−1
where Xit is a vector describing the user i and update t and β
is the model parameter vector.
MLE training based on “random bucket” data that has a
reduced serving bias. Big data and high dimensionality lead to
distributed training procedures e.g., ADMM, Spark.

9. Features
Xit = (Xi , Xt, Xik, Xij , Xijk, Xio)
Xi : viewer features
e.g., profile information
Xt: update features
e.g., actor type, verb type, etc.
Xik viewer-activity type features
“viewer-activity type affinity” (computed offline)
Xij viewer actor features
“viewer-actor affinity” (computed offline)
Xijk viewer-actor-activity type features
“contextual viewer-actor affinity” (partially computed offline)
Xio viewer object features
e.g., viewer-object content features

10. Affinity Features
Double or triple interaction features such as Xij or Xijk are hard to
scale due to the explosion of feature dimensionality.
Solution: Build a separate model for viewer-actor affinity αij ,
viewer-activity type affinity αit, and contextual viewer-actor
affinity αijk.
Affinity models are trained via a separate pipeline and stored
offline
At serve time, relevant affinity scores are loaded from data
store and are used as features in the CTR prediction model
Decouples a very complex problem into several simpler
problems and at the same time separates heavy computation
in offline and fast loading in online.

11. Joint Feed Effects
Some joint feed effects are hard to model using a “pointwise”
model such as a CTR model.
For example, we observe that engagement drops for non-diverse
feeds

12. Joint Feed Effects
We also observe that engagement drops for previously seen
updates.

13. Reranking
Ranking based on CTR scoring model cannot take into account
entire feed. Reranker manipulates the CTR ranked list to have a
better ranking and lower negative “joint feed effects”.
Reranking stages include
ensuring actor diversity (within session)
ensuring actor diversity (cross session)
ensuring update type diversity
impression discounting
additional “secret sauce”

14. Open Problems
AB tests for social networks
Explore/Exploit with minimal negative effect on member’s
experience
Likes, comments, shares create viral updates that propagate in
a network. How do we measure impact of a social gesture on
the eco-system
Tradeoff between optimizing for consumers and producers
Tradeoff between value to members and value to eco-system

15. Thank You
Additional details can be found in the Feed Relevance KDD 2014
(Activity Ranking in LinkedIn Feed) and KDD 2015 (Personalizing
LinkedIn Feed) papers on feed relevance.