Relevancy

Ranking
While implementing Learn-to-Rank systems, Metarank is designed to be a satellite secondary reranking system. It assumes that there exists another service which generates candidates for the reranking process:
in search: Elasticsearch or SOLR
in recommendations: output of spark mmlib ALS recommender
in ecommerce: inventory database
Most of these primary sources of input may also have a per-item score: how much this item is matching the original query:
BM25 or TF/IDF score in search
cosine difference between embeddings in recommendations
Metarank ranking event schema allows adding a per-item fields, which can be used for relevance score, see the example:
{
  "event": "ranking",
  "id": "81f46c34-a4bb-469c-8708-f8127cd67d27",
  "timestamp": "1599391467000",
  "user": "user1",
  "session": "session1",
  "fields": [
      {"name": "query", "value": "cat"},
      {"name": "source", "value": "search"}
  ],
  "items": [
    {"id": "item3", "fields": [{"name": "relevancy", "value": 2.0}]},
    {"id": "item1", "fields": [{"name": "relevancy", "value": 1.0}]},
    {"id": "item2", "fields": [{"name": "relevancy", "value": 0.1}]} 
  ]
}
This per-item "relevancy" field is the one holding information about score from the upstream ranking system, like BM25 score.
Metarank <= 0.5.11 included now deprecated relevancy extractor. With Metarank 0.5.12+ you can use regular number extractor for this case:
- name: relevancy
  type: number
  scope: item
  field: item.relevancy
Multiple retrievers
As there can be multiple per-item fields in the ranking event, it means that it's also possible to have multiple first-level relevancy signals. For example, when you have a hybrid search application with two retrievals:
Elasticsearch/OS/Solr for regular term search, giving you a per-document BM25 score.
PineCone/Vespa/QDrant/etc. vector search engine, doing a k-NN lookup over neural-based query embedding, giving you cosine similarity score.
The app retrieves top-N documents from both sources, and then merges them together in a single list. Some documents may come from both retrievers, and some - only from one.
Then your ranking event may look like:
{
  "event": "ranking",
  "id": "81f46c34-a4bb-469c-8708-f8127cd67d27",
  "timestamp": "1599391467000",
  "user": "user1",
  "session": "session1",
  "fields": [
      {"name": "query", "value": "cat"}
  ],
  "items": [
    {"id": "item3", "fields": [{"name": "bm25", "value": 2.0}]},
    {"id": "item1", "fields": [
      {"name": "bm25", "value": 1.0}, 
      {"name": "cos", "value":  0.75}
    ]},
    {"id": "item2", "fields": [{"name": "cos", "value": 0.02}]} 
  ]
}
In this case, item1 retrieved by both search engines (so there are two relevancy factors: bm25 and cos), item2 is coming only from vector search, and item3 only from term search engine.
Not supplying one of relevancy scores is possible, as underlying ML model implementations both support handling missing values (see docs for XGBoost and LightGBM on how it's done under the hood).
Position
A bias elimination technique based on a paper PAL: a position-bias aware learning framework for CTR prediction in live recommender systems.
on offline training, the feature value equals to the item position in the ranking
on online inference, it is equals to a constant position value for all items.
The main idea of such approach is that the underlying ML model will learn the impact of position on ranking, but then, setting all items position factors to the same constant value, you make it look like from the model point-of-view that all items are located on the same position. So position has constant impact on the ranking for all the items.
To configure this feature extractor, use the following YAML snippet:
- type: position
  name: position
  position: 5
To choose the best position value:
Start with a value around middle of your average ranking length. So if you present 20 items, set it to 10. Usually it's already a good number for most of the cases.
Do a grid search of the best value around it with metarank standalone. Select the best position based on offline NDCG value.
Generic
Scalars
Last modified 4mo ago