When the number of cold start items is large, for example, a large number of new short videos are uploaded to a short video platform every day and a large number of new items are launched on an e-commerce platform every day, it is inefficient to distribute traffic to all cold start items in order to explore their potential. In this case, candidate generation strategies are required.
Random candidate generation
Random candidate generation is fair but inefficient and can be used as a benchmark strategy.
Rule-based candidate generation
This strategy generate candidate items based on service features. For example, a dating platform can recommend candidates in the same city based on geographical locations of users.
Content preference-based candidate generation
This strategy collects statistics on or generate models of user preferences for item content or attributes, and then generates candidate items based on the preferred content and attributes.
Model-based candidate generation
This strategy requires a model that is applicable to cold start scenarios, such as the DropoutNet.
For information about the process of candidate generation by using the DropoutNet model, see