Abstract: There are a subset of members at LinkedIn who are receiving a large amount of irrelevant connection requests. Once they have accepted these requests, their network becomes overrun with irrelevant feed updates and notifications. At LinkedIn, we have seen this both in the data and feedback from our members. We have looked deeply into healthy networks and how we can help our members achieve them.
One approach is being smarter at recommending potential connections. When AI models are built to recommend member connections, the primary objective is to maximize the probability of forming a connection. However, this leads to some members receiving too many invitations. We addressed this issue by accounting for the recipient's invitation volume through a decay model, which reduces the appearance of members who have received too many invitations from other members’ recommended list.
Evaluating the impact of the decay model is a challenging problem with both sender and receiver side impact as well as a long term network effects. Sender impact can be measured directly through AB testing. Recipient side impact can be measured through an attribution framework. For long term network effects, we’ve designed special experiments to understand this more.
The learning objectives are:
Consider the experience of both sender and receiver in modeling the social network
Understand how the density of connection requests impacts the LinkedIn experience
Measure the impact of recommendations from both sender and receiver sides holistically
Bio: Divya is a Senior Machine Learning Engineer at LinkedIn. She is working on developing models recommending people for you to connect with on LinkedIn. Her academic and industry experience is in the areas of Recommender Systems and Software Engineering. She earned her Masters from Carnegie Mellon University.