2005/020 - Probabilistic latent clustering of device usage
- Jean-Marc Andreoli,Guillaume Bouchard
IDA 2005, the 6th International Symposium on Intelligent Data Analysis, Madrid, Spain, September 8-10, 2005.
We investigate an application of probabilistic latent semantics to the problem of device usage analysis in an infrastructure in which multiple users have access to a shared pool of devices capable of delivering different kinds of service and service levels. Each invocation of a service by a user, called a job, is assumed to be logged simply as a co-occurence of the identifier of the user and that of the device used. The data is best modeled by assuming that multiple latent variables (instead of a single one as in traditional PLSA) satisfying different types of constraints explain the observed variables of a job. We discuss the application of our model to a printing infrastructure in an office environment. Here, a job is assumed to be generated by two independent factors: the location of the user and the service class of the job (eg. black and white vs. colour, high volume vs. low volume etc.).