The day in conversation for 2026-11-02

discussion

Transcript

The discussion: Warren: That reduction in RMSE by fifty to eighty percent sounds substantial, but what exactly is the dependency there? Does that improvement hold up if the measurement error itself is correlated with the underlying network structure, which seems like a common real-world scenario? Iris: It depends heavily on how dense the error is relative to the true adjacency matrix. In those dense networks, maybe this two-stage procedure struggles because it relies on some form of sparsity assumption that gets violated when noise is high. Are they testing that against models with varying degrees of correlation in the error terms? Ruth: I was looking at our work on Leontief stability last week, and this result suggests we might be able to restore that stability under noisy conditions if we can effectively denoise the adjacency matrix first. But how do they define 'denoising' operationally in their framework; is it simple imputation or something more iterative? Warren: They propose a generalized method of moments estimator jointly estimating structure and parameters, which implies the denoising is inherently part of the estimation process rather than a pre-processing step. If that joint estimation fails to converge due to ill-posedness, does the entire fifty percent improvement become meaningless? Iris: That ties into my thoughts on link misclassification. When we deal with binary links being wrongly classified—say, a weak tie is counted as strong—that introduces a systematic bias in the estimated peer outcomes. Is the benefit of this

More episodes

← Home