Efficient Network Automatic Relevance Determination

We propose Network Automatic Relevance Determination (NARD), an extension of ARD for linearly probabilistic models, to simultaneously model sparse relationships between inputs X ∈ mathbb Rd × N and outputs Y ∈ mathbb Rm × N, while capturing the correlation structure among the Y. NARD employs a matrix normal prior which contains a sparsity-inducing parameter to identify and discard irrelevant features, thereby promoting sparsity in the model. Algorithmically, it iteratively updates both the precision matrix and the relationship between Y and the refined inputs. To mitigate the computational inefficiencies of the mathcal O(m3 + d3) cost per iteration, we introduce Sequential NARD, which evaluates features sequentially, and a Surrogate Function Method, leveraging an efficient approximation of the marginal likelihood and simplifying the calculation of determinant and inverse of an intermediate matrix. Combining the Sequential update with the Surrogate Function method further reduces computational costs. The computational complexity per iteration for these three methods is reduced to mathcal O(m3+p3), mathcal O(m3 + d2), mathcal O(m3+p2), respectively, where p ll d is the final number of features in the model. Our methods demonstrate significant improvements in computational efficiency with comparable performance on both synthetic and real-world datasets.