Evolutionary game theoretical on-line event detection over tweet streams

Current Online Social Networks represent a means for the continuous generation and distribution of information, which is slightly changed when moving from a user to another during the traversing of the network. Such an amount of information can overcome the capacity of a single user to manage it, so it would be useful to reduce it so that the user is able to have a summary of the information flowing the network. To this aim, it is of crucial importance to detect events within such an information stream, composing of the most representative words containing in each information instance, representing the event described by the set of tweet categorized together. There is a vast literature on off-line event detection on data-sets acquired from online social networks, but a similar solid set of approaches is missing if the detection has to be done on-line, which is demanding by the current applications. The driving idea described in this paper is to realize on-line clustering of tweets by leveraging on evolutionary game theory and the replicator dynamics, which have been used with success in many classification problems and/or multiobjective optimizations. We have adapted and enhanced a evolutionary clustering from the literature to meet the needs of on-line tweet clustering. Such a solution has been implemented according to the Kappa architectural model and assessed against state-of-the art approaches showing higher values of topic and keyword recall on two realistic data-sets.