Subspace-dependent adjacency matrix design via discrete-continuous optimization

We present a method for adjacency matrix design by applying both discrete and continuous optimization techniques, which are powered by energy minimization and gradient descent respectively. Most of the subspace clustering methods design the adjacency matrix by adopting a common calculation rule over the entire data set, which is under the assumption that a same noise model is shared between points. Therefore, these methods tend to fail when the points contain multi-source noises. To relax this limitation, we introduce a method that finds the neighborhoods by considering discrete-continuous optimization. Our method outperforms competitive methods on both synthetic and real-world data.