Reduce-Then-Predict or Simultaneous Reduce-and-Predict? Data-Driven Sparse Modeling for Improving R&D Efficiency
针对PCB设计公司研发数据有限、噪声大等问题，比较了先降维再预测（SPCA-LASSO）与同时降维预测（SPCR）两种稀疏建模方法，发现前者预测误差降低22-41%，且结果更稳定可解释。
Efficient Research and Development (R&D) workflows are critical in industries where early-stage results influence downstream outcomes. This study develops a predictive model to enhance R&D efficiency for a leading integrated device manufacturer (IDM) specializing in Printed Circuit Board (PCB) design. To address challenges of limited data, noise and collinearity, we apply Sparse Principal Component Analysis (SPCA) to simplify simulation data, followed by LASSO regression to predict later-stage physical testing performance. Our SPCA-LASSO model reduces prediction errors by 22–41% compared to direct LASSO regression while offering interpretable insights for engineers. In contrast, Sparse Principal Component Regression (SPCR), which integrates dimension reduction and prediction, yields higher errors and unstable factor loadings. This empirical comparison between reduce-then-predict and simultaneous reduce-and-predict approaches contributes to sparse modeling and engineering analytics, offering actionable insights for improving sequential R&D processes across high-tech industries, software engineering, construction and other sectors where early performance predictions are critical.