In classification learning of artificial stimuli, participants learn the perfectly diagnostic dimension better than the partially diagnostic dimensions. Also, there is a strong preference for a unidimensional categorization based on the perfectly diagnostic dimension. In a different experimental procedure, called array-based classification task, participants do not exhibit a preference for a unidimensional categorization. In Experiment 1, we replicate the above results. In Experiment 2, we show that when participants learn the partially diagnostic features through repeated testing, there is a decrease in unidimensional categorization. We use Bayesian modeling to show that only those participants who learned the diagnosticity of a dimension with a high level of accuracy ([greather than or equal to]75%) used the dimension for categorization. Our results show that whenever accurate knowledge about feature diagnosticities is available, there is a lesser preference for unidimensional categorization. Our results provide a possible explanation for the preference of unidimensional categorization in classification and observation learning.