Boost by Popularity: Using Engagement Data in Document Ranking
Doug Heitkamp, Mark Tolram, Julien Vaillant · SSRN Electronic Journal · 2019
Customer feedback, validated by historical engagement data, showed that navigational queries consistently under-performed. By analysing the click curve distributions for these queries we were able to hypothesise that including document popularity as a relevance signal could address this particular customer need and positively impact other documents frequently engaged with below the fold. For this piece of work, we created a Python algorithm to generate query-document pairs scores, and were able to introduce them successfully in our search engine algorithm. This allows us to add a new signal to our relevance calculation. We have since resolved many of our customer issues and report a significant increase in DCG overall.