Modern computer applications, from business decision support to scientific data analysis, utilize data visualization tools to support exploratory activities. Visual exploration tools typically do not scale well when applied to huge data sets, partially because being interactive necessitates real-time responses. However, we observe that interactive visual explorations exhibit several properties that can be exploited for data access optimization, including locality of exploration, contiguous queries, and significant delays between user operations. We thus apply semantic caching of active query sets on the client side to exploit some of the above characteristics. We also introduce several prefetching strategies, each exploiting characteristics of our visual exploration environment. We have incorporated caching and prefetching strategies into XmdvTool, a public-domain tool for visual exploration of multivariate data sets. Experimental studies using synthetic as well as real user traces are conducted. Our results demonstrate that these proposed optimization techniques achieve significant performance improvements in our exploratory analysis system.