LinkedIn Recommendation

It's really good to see the practical recommendation case from industry via MOOCs (Introduction to Recommender Systems).

• Hybrid

Anmol Bhasin introduces the LinkedIn recsys overview and the system has various recsys for different tasks (job, people, article/newsfeed etc.) with hybrid strategies. The top-k recommendations from different recsys are mixed and optimized for a particular P (click rate, revenue, connection rate etc.) where constraints can denote such as certain portion of clicks to LinkedIn pages and others to external pages.

• Magic is not in a model, but features

Data is tremendous, but take into account the right features to models is more important. For instance, location is very important, but sometimes has been ignored by recsys. Students care less about locations while professionals don't.

• Use extensively crowdsourcing for learning something from different context data

• Different features and models for different populations (tuning regarding to the segments of the population)

• Evaluation offline an online
• Evaluation metrics may/may not correlate to the actual effects (interaction in the two settings is different, e.g., if show many times of an item online, then the user might click it - which cannot be captured in offline settings)
• A/B testing (sometimes called split testing) is comparing two versions of a web page to see which one performs better: if the evaluation metrics you were testing is promising, then do the A/B testing with segments of people by different settings of traffic (recruiters, talents).