Business rules in recommendation algorithms define how to behave with the items in a potential recommendation set in case the rule is applied. This can, for example, filter out items based on a certain criteria or change a recommendation score and thus the ranking of items.
Business rules can generally be applied to all algorithms. However, some rules are user-specific, meaning that some information about the user (most often, user history), is required and the algorithm must produce personalised recommendations for the rule to work.
Currently, there are 4 generic business rules implemented in PEACH.
This is a simple rule which allows to filter out certain programs from recommendation sets. The blacklisted programs are still used for the algorithm model training and therefore affect the modelling of a user's taste but do not appear in recommendations.
The program lifetime rule allows to rescore an episode based on its age since publication and a computed relevance period of the program it belongs to. If an episode is too old (i.e. less relevant), it will be downscored and therefore will have a lower chance to be recommended. Relevance scores are computed automatically based on program watch/listen events and no additional metadata is required.
The implicit favourites rule collects programs which users frequently consume and marks a given program as an "implicit favourite" if the user consumed more than a certain threshold number of episodes of that program. New episodes that belong to a program which is an implicit favourite are upscored and therefore more likely to be recommended.
Limit of episodes per program
This simple rule sets an upper threshold for the number of episodes belonging to the same program to appear in any given recommendation set, potentially increasing the diversity of programs covered by recommendations.
Applying business rules
Since typically several business rules are applied at once in a recommendation endpoint, a standard way for applying them has been introduced.