Search Ranking Models (Learning to Rank - LTR)

Has anyone explored search ranking, normally called Learning to Rank (LTR), either using traditional or deep learning techniques? I’m just trying to get started in this area and am having difficulty finding good practical advice on how to implement and train these types of models. I’d love it if anyone could recommend a course/book/blog that follows a similar teaching style as :slight_smile:

The tensorflow team announced a LTR library that I’m excited to try soon.

I’m quite interested in this topic too. Airbnb recently published a very interesting paper documenting their journey to improve on their previous production model ― based on a GBDT model and a lot of heavy feature engineering ― in order to break out of that plateau in terms of improvements of their online metric (number of bookings).

I am thinking about writing LTR (mostly Lambda Rank) in Fastai. I saw the paper from Airbnb as well. Anybody interesting in collaboration?

I develop Search and Recommendation systems for work and I have a good pulse on the field so I can answer your questions.

Currently, like most fields, Information Retrieval (IR) is being changed by deep learning. Although boosting methods (LambdaMART, etc) still prove as powerful baselines, DL methods are become more popular and powerful. For example, YouTubes paper on their deep search and recommendation system has driven a lot of innovation in industry: .

I would suggest trying out the TF Ranking library and try to implement RankNet which was one of the first ranking algos to use NNs (it drove Bing search for a while I’ve read…).

Yes, also the Airbnb paper @iamgianluca mentioned is very good and has a lot of practical techniques for industry: weighted item2vec, clustering items and users in Matrix Factorization to reduce sparsity, LTR relevance labels derived from log feedback, etc…

1 Like

What about creating a page similar to this where we could collect useful resources for anyone interested in DL for ranking problem?