Fractal trees learn decomposable models with a bounded clique size.

Fractal trees have a computational complexity of O(k²⋅n²⋅N)$\mathcal{O}(k^2\cdot n^2\cdot N)$, in the worst case.

We propose a prune-and-graft operator with the same computational complexity.

All the provided procedures are modular and can be easily parallelized.

The proposed algorithms have obtained competitive experimental results.