Novel methodology for obtaining design structure matrices using network identification

What the math gives to ML

The paper presents a constructive method for inferring a design structure matrix, equivalent to a directed adjacency matrix, from observed dynamical data instead of relying only on expert knowledge. Its transferable mechanism is sparse network identification: estimate directed couplings in a state-space or dynamic-network model, then convert statistically supported couplings into a weighted or binary dependency graph. In neural networks, this can learn sparse block connectivity between hidden groups, recurrent state variables, experts, or layers, while preserving the identified interaction structure during training. The key falsifiable prediction is that thresholding identified couplings should produce a sharp precision-recall transition near the estimation-noise floor, while computational cost decreases approximately in proportion to the retained edge fraction.