Bayesian Regression for Interpretable Network Dimensioning
Shrihari Vasudevan, Sleeba Paul Puthenpurakel, Marcial Gutierrez, MJ Prasath · 2023
This paper proposes a solution to learn interpretable regression models for the Telecommunications industry's network dimensioning problem. Domain knowledge in the form of expected model coefficients and desirable trends among them are incorporated into the modeling process, using an approach based on Bayesian Regression. This not only enables intuitive models for domain experts but simultaneously addresses known multicollinearity issues in the underlying data-set. The models thus trained are nearly as-performant as the best theoretical model for the data-set. The proposed solution has been deployed in production. Learnings from this paper are transferable across Telecommunications problem-contexts and application domains.