CPrefMiner: A Bayesian Miner of Conditional Preferences

Customizing database queries through the use of user preferences is a research topic that has been raising a lot of interest within the database community in recent years. Such preferences are used for sorting and selecting the best tuples, those which most fulfill the user wishes. A topic of interest within this context is the elicitation of preferences, consisting of methods to enable the user to inform his choice on pairs of objects belonging to a database. Depending on the size of the database, this task may require a great effort from the user, and consequently may discourage him/her to use the system. In this paper, we propose a first step towards the design and implementation of an automatic tool for inferring preferences from a given sample of user preferences. Our method CPrefMiner we propose is based on the framework of Bayesian Networks and aims at mining a special kind of preferences, the conditional preferences. The two main learning tasks accomplished by CPrefMiner are: (1) Learning the graph underlying the conditional preference network; (2) Learning the preference probability tables associated with each node of the graph. This paper focus on the first task.