Ranking

Choose ranking models for two different jobs: retain relevant candidates from a large set, then order the smaller set carefully. This lesson compares pointwise scoring with pairwise learning and introduces LambdaRank and LambdaMART.

Objective

The model should put relevant results near the top of the SERP. Evaluate the top positions explicitly, because the design places more importance on the results users encounter first.

Learning to rank (LTR): A class of techniques that applies supervised machine learning (ML) to solve ranking problems. The pointwise and pairwise techniques that we will apply fall under this class.

Stage-wise approach

As discussed in the architectural components section, the number of documents matching a single query can be very large. So, for a large scale search engine, it makes sense to adopt a multi-layer funnel approach. The top layer of the funnel looks at a large number of documents and uses simpler and faster algorithms for ranking. The bottom layer ranks a small number of documents with complex machine-learned models.

Let’s assume here that you will use two layers for ranking with one-hundred thousand and five-hundred documents ranked at each layer, as shown below. Though, the choice of layers and documents ranked on each layer depends highly on the available capacity.