We present a method for automatically learning an effective strategy for clustering variables for the Octagon analysis from a given codebase. This learned strategy works as a preprocessor of Octagon. Given a program to be analyzed, the strategy is first applied to the program and clusters variables in it. We then run a partial variant of the Octagon analysis that tracks relationships among variables within the same cluster, but not across different clusters. The notable aspect of our learning method is that although the method is based on supervised learning, it does not require manually-labeled data. The method does not ask human to indicate which pairs of program variables in the given codebase should be tracked. Instead it uses the impact pre-analysis for Octagon from our previous work and automatically labels variable pairs in the codebase as positive or negative. We implemented our method on top of a static buffer-overflow detector for C programs and tested it against open source benchmarks. Our experiments show that the partial Octagon analysis with the learned strategy scales up to 100KLOC and is 33x faster than the one with the impact pre-analysis (which itself is significantly faster than the original Octagon analysis), while increasing false alarms by only 2%. The general idea behind our methodis applicable to other types of static analyses as well. We demonstrate that our method is also effective to learn a strategy for context-sensitivity of interval analysis.