Upleveling Collaborative Articles’ Match Game

A year ago we launched collaborative articles with the vision of making LinkedIn the go-to destination for all your work-related questions. Collaborative articles are AI-powered conversation starters published by LinkedIn with expertise, knowledge, and insights shared from the LinkedIn community.

In November 2023, we published an overview of how we developed this system and what we learned six months in. In this post, we will focus on how we incorporated feedback from our members and evolved our systems to improve how we match articles with members who demonstrate expertise in certain areas.

As mentioned in our earlier blog post, a crucial component of this system is identifying genuine experts, which turned out to be more difficult than expected. Despite having numerous direct and indirect indicators to measure a member’s proficiency in a skill, there can be a significant amount of interference. Some of these indicators have limited coverage, while others lack accuracy. In the end, the indicators that performed reasonably well were a mix of explicit skills – skills listed on profiles, skill endorsements from others, recent job titles – and implicit skills, which are deduced from various signals at LinkedIn – recent hires for job postings or a member’s self-assessment during job applications, to name a few. 

Furthermore, we incorporated insights, such as the probability of a member contributing original ideas based on their previous content shared on LinkedIn. This gave us a reasonable starting point for identifying the most knowledgeable individuals. Once they began sharing their thoughts, we refined our process and evaluated how engaging and valuable their opinions were to the community. This is where the transformation occurred – converting a combination of raw indicators into practical insights enabled us to better match member experts with articles where they could contribute their viewpoints, responses, and insights.

Using Embedding Based Retrieval For Improved Matching

One of the most interesting segments of our work thus far is how to build the technology that powers our ability to make matches between topics and contributors. Once we’d identified the right members, we needed to match them to the best questions they’d be excited to answer. 

Based on feedback from our members through our evaluation mechanisms, we focused our efforts on our matching capabilities between articles and member experts. One of the new methods we use is embedding-based retrieval (EBR). This method generates embeddings for both members and articles in the same semantic space and uses an approximate nearest neighbor search in that space to generate the best article matches for contributors.

Skills served as the keystone to connect our member experts and the articles. To make this possible, we first created embedding vectors for skills, storing results in a skill-id-to-embedding map. Then, proficient skills for each member were determined. Embeddings from the skill-id-to-embedding map were averaged for members with multiple skills, leading to a member-to-skill embedding map. 

Wanting to do the same for articles, we crafted embedding vectors for articles, and these vectors served as EBR queries in the skill-id-to-embedding index, revealing article-relevant skills. Skill vectors were averaged, resulting in an article-to-skills map. 

Once we had a map of article embeddings, whenever we needed the member expert matches for the most relevant questions, we fed in the member-skill embeddings as queries to this index. Using an approximate nearest neighbor algorithm, we could then retrieve the top 10 documents that were closest to the query in the skill-embedding space, and as a result, we received a fourth map from each member to a list of articles with the articles ordered by their distance in the above-described semantic space.

These EBR-based article candidates are then ranked along with other skill-based matched candidate sources using a supervised AI model to generate the final list of the most relevant ranked articles we think a member would want to contribute to. 

We’re Still Learning and Growing

It has only been 12 months since the introduction of collaborative articles, and we have already witnessed a 4x increase in weekly member contributions quarter-over-quarter and over 10 million contributions from the LinkedIn community. As engineers, there is no greater satisfaction than providing individuals with a robust platform for expression on a grand scale. With the rapid advancement of AI and our continuous learning as engineers, we remain committed to incorporating features that translate into increased value for our members.