Ranking is widely used in product search and product recommendation world. It is widely employed for providing meaningful suggestive results to end-users ordered by relevance.
Ranking in recommender system is widely recognized as the primary way to order the items within recommendation system to provide the users relevant and meaningful suggestions. The purpose here is to recommend items to users with the intent of making it click or progress to the lower funnel stages.
Ranking approaches
Ranking approaches can be divided into three types – pointwise, pairwise and listwise.
Pointwise approach
Pointwise approach treats the ranking problem as regression or classification. Each item is assigned a relevance score independently. Examples include linear/logistic regression, XGBoost classifier, Neural Networks.
Pairwise approach
Pairwise approach focuses on correctly ordering pairs of items. It transforms the ranking into the binary classification of item pairs.
The difference between point wise and pairwise approach is that point wise approach doesn’t consider the relative ordering between items while pairwise consider the relative ordering between items.
Examples include RankNet, LambdaRank, SVMRank.
Listwise approach
Listwise approach consider the entire list of items to be ranked at once. It optimizes ranking metrics directly over the entire list. Examples include ListNet, List MLE and LambdaMART (combines pairwise and listwise approaches).