A series of improvements to LinkedIn’s search system will allow publications to deliver faster, more relevant results within a streamlined user interface.
LinkedIn details the development journey of the new system architecture while explaining the complexity of its legacy system.
Post search results were previously presented by two indexes – one for posts in the LinkedIn main feed and one for articles.
The complex nature made it difficult to build on, so LinkedIn decided to separate the two franchises. LinkedIn reveals the entire process in excruciating detail in a new blog post.
Here it is summarized in an infographic:
Most of the information shared by LinkedIn is geared towards software engineers. In this post, I’ll go over the technical details and explain what the changes mean for regular users on LinkedIn.
Search published on LinkedIn provides faster and more relevant results
To make search results more relevant, LinkedIn set out to create a system that takes into account the following aspects:
- Relevance of the function to the query
- Post quality
- personalization
- User intent
- link
- Freshness / novelty
LinkedIn says its new system should also deliver results from a variety of sources.
LinkedIn leverages several machine learning techniques to meet the searcher’s expectations of achieving these goals in terms of relevance and diversity.
What’s more, LinkedIn crowdsourced human ratings of search results and leveraged the data to ensure that its new system meets a certain quality threshold.
LinkedIn notes that its crowdsourced human annotation data also provides valuable training data for improving the ranking of results.
new system vs. Old system
Powered by machine learning, the new LinkedIn system improves on the old system in the following ways:
- Relevance: Allows personalization by leveraging deeper, real-time signals of members’ intentions, interests, and affiliations.
- diversity: Increases detection of potentially viral content for popular search queries and reduces duplication of similar content.
- classification: Uses job-related metadata from the index to improve the ranking of posts when mixed with other types of results.
- Mobility: It has a new user interface that allows people to search for posts by a specific author, posts that match quoted queries, recently viewed posts, and more.
The data shows that the new system is better
LinkedIn says that the search results provided by its new system have increased user satisfaction, which is reflected in the 20% increase in positive feedback.
Results closely related to a user’s query, which LinkedIn identifies as “relevant results,” resulted in an overall improvement in click-through rate. more than 10%.
The wide variety of posts from within the researcher’s social network, geographic location, and preferred language generated a file 20% increase In correspondence within the researcher network.
The time it takes to provide search results has been reduced by ~62 ms for android devices, ~34 ms for iOS and ~30 ms for web browsers.
future improvements
LinkedIn shares how it will continue to improve post search results. Future updates will include:
- Implement natural language processing to understand the semantic meaning of queries.
- Highlight the latest results for queries on popular topics, and reduce the feedback loop from hours to minutes.
- Extend document understanding capabilities to work with multimedia content such as images, short video clips, and audio.
See the full LinkedIn blog post for all the technical details behind these changes.
Source: linkedin
Featured image: / shutterstock
