Community-Focused Feed Optimization

June 25, 2019

Co-authors: Ali Mohamed and Zheng Li


LinkedIn’s feed stands at the center of building global professional knowledge-sharing communities for our members. Members talk about their career stories, job openings, and ideas in a variety of formats, including links, video, text, images, documents, and long-form articles. Members participate in conversations in two distinct roles: as content creators who share posts, and as feed viewers who read those posts and respond to them through reactions, comments, or reshares. By helping members actively participate in those professional conversations, we are fulfilling LinkedIn’s mission to connect the world’s professionals to make them more productive and successful. 

This post focuses on one important aspect of  the machine learning at Linkedin’s feed: candidate selection. Before final fine feed ranking, a personalized candidate generation algorithm is applied to tens of thousands of feed updates to select a diverse and unique candidate pool. We describe the machine learning models applied in candidate generation and the infrastructure capabilities that support accurate and agile model iterations.

Overview of Feed at LinkedIn

At the heart of the feed sits a machine learning algorithm that works to identify the best conversations for our members. In a fraction of a second, the algorithm scores tens of thousands of posts and ranks the most relevant at the top of the feed. In order to operate at this scale and speed, LinkedIn’s feed has a two-pass architecture. The first pass rankers (FPR) create a preliminary candidate selection from their inventories based on predicted relevance to the feed viewer. Examples include updates from your network, job recommendations, and sponsored updates. A second pass ranker (SPR) then combines and scores the output from all first pass rankers. The SPR creates a single, personalized ranked list. FollowFeed is the dominant FPR that serves feed updates from your network. More than 80% of feed updates are coming from FollowFeed, and those updates contribute to more than 95% of members’ conversations. Through these conversations, active communities are formed and strengthened.

  • two-pass-ranking-architecture

Two-pass ranking architecture for LinkedIn’s homepage feed

At LinkedIn’s scale, the main technical challenge is to find the right balance between infrastructure impact and multi-objective optimization using comprehensive machine learning algorithms. Those objectives include members’ likelihood to view updates, their likelihood to participate in conversations, and providing timely reactions to content creators. There are hundreds of machine learning features we use to compute these objectives. We want to optimize these objectives continuously and accurately while satisfying the low latency requirements of our infrastructure footprints.

Our teams tackled this problem through a joint project amongst the Feed Artificial Intelligence (AI), Feed Infrastructure, and Ranking Infrastructure teams. As we ramp this project globally, we would like to share more details about our technical solutions.

  1. We refreshed the machine learning software stack in FollowFeed, leveraging the latest productive machine learning technologies. Through a ranking engine upgrade and model deployment technology, we enabled frequent and agile updating of machine learning models. We also added accurate tracking of machine learning features in FollowFeed, which helps us guarantee data and model consistency across training and serving. Moreover, we developed tools to inspect machine learning algorithm complexity at serving time.

  2. With minimal additional complexity, we have rebuilt our machine learning model for candidate selection from scratch with new objectives and different algorithms. As part of this, we introduced the prediction of professional conversation contribution into our model to capture the community-building aspect of each feed update. Instead of multiple logistic regressions with manual interactions, we’ve used a single XGBoost tree ensemble to trim down the complexity. Additionally, we’ve considered timely feedback to content creators in our model and make sure all members have a chance to feel heard. All of these things are done with minimal infrastructure capacity addition.

In summary, this project builds the engineering capabilities for us to iterate comprehensive machine learning models at the candidate generation stage, and we’ve leveraged these capabilities to deliver performant, multi-objective optimization models. At LinkedIn’s scale, these solutions will help bring economic opportunity to the global workforce through more relevant conversations with their professional communities.

Performant AI infrastructure with agility

FollowFeed, LinkedIn’s powerful feed indexing and serving system, has now been equipped with advanced machine learning capabilities. The initial design had accommodated the ranking needs for feed, but the field of machine learning has advanced tremendously in the past five years since the original FollowFeed design. During this project, we boosted the agility and productivity of machine learning in FollowFeed by adopting the latest machine learning inference and model deployment technologies. Such infrastructure enhancements enable the modeling capability described later in this blog post.

  • FollowFeed-architecture

FollowFeed architecture

We have updated FollowFeed’s ranking engine to Quasar. Quasar, as part of LinkedIn’s Pro-ML technology, transforms machine learning features and inferences the machine learning model at query time. As a high-performance, multi-threaded ranking engine, Quasar not only optimizes for infrastructure system efficiency but also machine learning productivity. Such productivity improvements have enabled:

  1. Cross-system leverage: We can easily port the latest machine learning models and transformers from the second pass layer to FollowFeed.

  2. Training and serving consistency: At offline training time, the same codebase is used to represent the model as at serving time. 

To reflect the rapid evolution of LinkedIn’s content ecosystem, machine learning models have to be constantly updated. We’ve built FollowFeed’s model deployment on top of LinkedIn’s Continuous Integration and Deployment (CICD) stack. Being a stateful system that indexes members’ past activities, FollowFeed presents a unique challenge in model deployment. We have to avoid calling external services to maintain high reliability and performance of index nodes where ranking is taking place. To optimize for such limits, we previously coupled ranking models with code, which leads to strong coupling of service deployment with model coefficient changes. To allow for easier model evolution, our solution is now a data pack-based deployment model, where we package the machine learning models in a separate code base, a.k.a. a multi-product. Those models are treated as a “data pack,” a deployable package consisting only of static files to be dropped into a specific location of production machines. Through such a design, model deployment can be easily managed by LinkedIn’s CICD system. Consequently, we’ve improved model deployment velocity from 3 days to 30 minutes.

In addition to ranking and model deployment, we designed and implemented advanced feature access and tracking in FollowFeed. As it scores thousands of documents per session, FollowFeed optimizes access to machine learning features needed by scoring models. Viewer features are passed down as part of the request without requiring an external call to be made. Features for feed updates are ingested, stored, and updated alongside these updates so that they are accessible locally on the index nodes for scoring. Given the well-known data inconsistency challenges between offline training and online scoring, we also added accurate tracking of machine learning features in FollowFeed. This guarantees data consistency between offline training data and online inference data. Aggregating these tracking events across the distributed index nodes presents a challenge. Even though results from the index nodes are aggregated in the broker layer, we do not want to gather the tracking data synchronously due to scalability and serving latency concerns. Our design overcomes these concerns by having individual index nodes and the broker node stream tracking events asynchronously to Kafka streams. A Samza job joins these Kafka streams and emits a comprehensive tracking event for the request. 

Equipped with advanced machine learning capabilities, it is much easier to develop performant machine learning models for FollowFeed with accurate features. Such agility will enable better candidate feed updates to be surfaced on LinkedIn’s homepage. Through actively updating the machine learning model, we will be able to give more power to existing and newly-minted content creators in LinkedIn’s ecosystem. We will also facilitate the professional community builders’ curation of their audience.

Optimizing for conversations at candidate generation

We’ve rebuilt the machine learning model at FollowFeed to optimize multiple objectives under given infrastructure constraints. The existing candidate selection is conducted through a logistic regression (LR) model that predicts the click probability. To facilitate professional conversations in LinkedIn’s feed, we have introduced “contribution” as an additional objective in the candidate selection model. Probability of contribution captures members’ intent to share, comment, or react to a particular feed update. The model also takes into account timely feedback to content creators, which is a clear signal for cultivating and retaining audience builders on LinkedIn.

To achieve these goals, our design evaluates candidate selection through recall instead of precision, combines multiple objective functions in a gradient boosting tree model, and trims down features and manual transformations.

Given that there is a comprehensive model at the second pass layer with more sophisticated computation per update, the job of candidate selection should be focused on generating a rich set of candidates. Clarifying such goals helps us evaluate our model much more effectively. Instead of precision, we use recall as an evaluation metric because it measures the overlap between the top K candidates generated by FollowFeed and the second pass ranker. K is very large for LinkedIn’s feed. To overcome the selection bias introduced by FollowFeed, we randomize its output for a small percent of traffic so that the second pass layer can be exposed to a representative range of candidates. This technique helps us approximate the second pass ranking at the candidate selection process. This technique helps us approximate what the second pass layer's ranking would be if the candidates were randomized and not pre-ranked by the first pass layer. Through various efforts outlined below, we have effectively doubled the recall percentage throughout the project.

  • recall-calculation-example

Examples of recall calculation

We use a logistic regression model with differential weights to incorporate additional contribution objectives in a single objective model. It helps reduce the parameter space and model complexity by half while effectively reproducing the effect of a multi-objective model.  Through a suite of tools to conduct CPU/network benchmarks, JVM model profiling, and feature importance analysis, we’ve simplified the model by replacing manual feature transformations with the gradient-boosted tree and trimming down excessive features. We utilized the same differential weighting technique to train the single tree ensemble. As gradient-boosted trees are powerful models themselves, we explored using a tree model alone, without logistic regression as the final layer, to predict the re-weighted objective. The list below shows various techniques we’ve tried, with different trade-offs in terms of infrastructure cost and online results. 

  • Baseline: The production model for FollowFeed for the past several years has been a logistic regression using manually-transformed features that predict the probability of clicks.

  • Tree + transformations + logistic regression: We added the contribution objective to our model. In addition to the existing transformed features, we also added XGBoost tree scorers as interaction features to the contribution-weighted-click logistic regression. This version performed much worse in terms of CPU utilization and 99th percentile latency, but online experiments showed strong lift in member engagement.

  • Tree + logistic regression: We removed manually-transformed features from the above implementation to reduce model complexity. Its infrastructure costs are still worse than baseline, but better than the previous implementation. We can keep similar engagement improvements.

  • Tree only: We use XGBoost tree scorer to calculate the contribution-weighted-click and remove the logistic regression component. We keep the majority of engagement improvements, This version has minimal additional overhead on CPU utilization and actually reduced tail ranking latency compared to the baseline. We’ve decided to ramp tree-only model to all members.

We’ve also adjusted the candidate selection results based on the freshness of the updates. This comes from observing a secondary creator-side effect while ramping an earlier version of the candidate selection algorithm. While viewers are less sensitive to the freshness, content creators do care about receiving prompt responses. By adjusting the freshness, viewers can provide creator feedback earlier and stay in active conversations.


The project’s models are already rolling out across LinkedIn. The feed updates are scored and ranked in Quasar. The machine learning models have been deployed through LinkedIn’s CICD system using FollowFeed’s data pack. Over the past few years, we’ve tried about 100 variations of the machine learning models and picked the best performing one through online A/B testing. We have seen a lot more members on LinkedIn's feed participating in professional conversations due to better candidate selection. We’ve also observed more reactions given to a broader range of creators so that they have additional ways to engage with content on the platform. We are looking forward to ramping this to all our members soon. In the meantime, we are continuing to innovate in our machine learning infrastructure and modeling technologies to bring the best conversations to members’ homepage feed.


The work described in this blog post is carried out by engineering talents across LinkedIn’s Feed AI, Feed Infra, and Ranking Infra teams, with strong support from Data Science, Relevance Explains, and SRE. In no particular order: Fei Chen, Greg Pendler, Hassan Khan, Ian Ackerman, Madhu Arun, Manas Somaiya, Parin Shah, Prasad Sriram, Prashant Saxena, Shihai He, Souvik Ghosh, Tim Jurka, and Ying Xuan.