Extending PSL with Fuzzy Quantifiers
Many interesting tasks in artificial intelligence require the ability to work with imperfect relational data. Examples include social network applications in which partial information about users and about their connections is available. For instance, we may have attribute information about users; we may know the age, gender, personality, likes and dislikes etc. of a user, but this knowledge is usually incomplete in the sense that we do not have all attribute values for all users. The knowledge can also be uncertain, for instance when the attribute values are not given directly by the user but instead they are inferred from user generated content. The relational aspect of social network data stems from the connections between the users, e.g., Facebook users are connected with their friends; Twitter users can follow one another; in Amazon, users can bookmark other interesting users; in Epinions users can trust other users or include them in their block list (i.e. distrust) etc. Again, the avail- able knowledge about the connections is typically incomplete and uncertain. An interesting task is to perform most probable explanation (MPE) inference over a given social network, i.e. based on (1) given attribute values about users and their connections and (2) some background knowledge about the domain, expressed as probabilistic rules in first- order logic, infer missing attribute values such that the given and the inferred values combined adhere to the background knowledge as well as possible. MPE inference is an important task in statistical relational learning (SRL).
Proceedings of StarAI2014 (4th International Workshop in Statistical Relational AI), workshop at AAAI2014 (28th Conference on Artificial Intelligence)