Facebook’s engineering team is sharing some of the technical aspects of Graph Search, including a discussion today about how it indexes and ranks data via its Unicorn framework — the basis of its new Graph Search engine.
In today’s blog post, Sriram Sankar talks about Facebook’s technical process for identifying and ranking Graph Search results — including how it uses click-thru rate, engagement and NDCG (normalized discounted cumulative gain) to “maximize searcher happiness.”
One of Facebook’s challenges is combining different data verticals across its graph, and then deciding how those different verticals should be ranked:
Scoring works on a single entity at a time and the score assigned is independent of the scores assigned to other entities. This can cause a result set to become very one-dimensional and offer a poor search experience, (for example, “photos of Facebook employees” may return too many photos of Mark Zuckerberg). Result set scoring offers yet another layer of filtering that looks at a number of entities together and returns a subset of these entities that are most interesting as a set (and not necessarily the highest scoring set of results).
If that’s sounding a bit technical, it is. The post is very high-level, but you might find a few nuggets of info in there — particularly in the “Scoring” section.
Also of note is further confirmation that Facebook is working to expand Graph Search so that user posts and comments are searchable. Facebook has said that’s in the works since Graph Search was announced in January.