Cold-Start Validation and Experiment Plan for LinkedIn Connection Suggestions

LinkedIn's "People You May Know" feature suggests other members a user might want to connect with. Suppose you have to test this feature but have no existing data to base the decision on. How would you test it, which metrics would you use and how would you decide whether it works? ```hint Data you can create Having no history does not mean having no evidence. Think about what you can learn before a launch and during a small one, and remember that every connection involves two members, not one. ``` ### Clarifying Questions - Is the feature brand new with no usage history at all, or does it exist and the missing data concerns a specific change to it? - What decision should the test inform: whether to launch, which suggestion logic to use, or where in the product to show it? - Which members matter most: new sign-ups with an empty network, or established members with many connections? - Can we run randomized experiments on live members, or must part of the decision be made before any launch? ### What a Strong Answer Covers - A clear goal and testable hypotheses for the feature, tied to the value members get from a relevant network. - Ways to produce evidence when no history exists, and how that evidence shapes the first version. - An experiment design that handles invitations crossing between test groups, plus duration, sizing and segments. - A metric set that rewards accepted, useful connections, with guardrails on the receiving side such as ignored invitations, reports and privacy complaints. - Decision criteria fixed before results arrive, and how the launch feeds later improvements. ### Follow-up Questions - Invitations sent rise sharply, but the acceptance rate falls. Do you ship? - A member complains that a suggestion exposed a private relationship. How do you respond in the product and in the test plan? - After a few weeks of data, how would you use it to improve the ranking of suggestions?