Efficiently estimating combinatorial t-way coverage in software testing remains a significant challenge due to the high-dimensional complexity of modern software systems and the inherent complexity of the problem. This article explores a novel method that uses matrix complexity metrics as features in machine learning algorithms to estimate the t-way coverage of random test sets. Based on an input parameter model of the system under test (SUT) we derive complexity metrics from the singular value decomposition of the matrix representation of the test set. This makes our approach independent of the SUT’s input space dimension and the size of the test set. Our approach provides a good estimation of the combinatorial t-way coverage while being faster and more scalable than a state of the art tool for the exact computation of combinatorial t-way coverage. Moreover, our experiments show a connection between the spectrum of singular values of a random test set and its t-way coverage. The use of complexity metrics as predictors in our machine learning pipeline adds a new dimension to the combinatorial testing domain, offering an additional tool for improving software testing processes.