Abstract: A method and system for transforming a multivariate data domain into a low-dimensional visual representation. Probabilistic models of the data domain are utilized, and at least one probabilistic model is used to produce at least one predictive distribution. The predictive distributions are used as inputs to the visualization process, where the multidimensional space is converted to a low-dimensional space. In this process data vectors are considered similar, for example, if the corresponding instances of a predictive distribution, conditioned with the variable value assignments found in the data vectors, are similar. Consequently, similarity is not defined directly using the physical properties of the data vectors, but indirectly through the probabilistic predictive model(s).