<p>We study a data-driven multi-item newsvendor problem with resource constraints. Demand of each item depends on exogenous features and a random shock. The objective is to obtain a data-driven ordering decision that minimizes the inventory cost. We assume that the firm has no prior knowledge on demand distributions, but has access to past demand samples and related feature information. We adopt local learning methods that approximate the objective function using weight sample average in which the weights can be computed by the <i>k</i>-nearest neighbors (<i>k</i>NN) and kernel regression methods. Such an approximation enables us to determine the order quantities directly from historical data. We then analytically derive the effectiveness of our proposed methods by showing their asymptotic optimality. That is, the ordering decisions derived from our methods and the corresponding costs converge to the optimal ones that are obtained by assuming known demand distributions. We also provide explicit performance bounds for our proposed methods in terms of the number of items, the feature dimensions and the sample size. Finally, numerical studies show several observations: (1) our methods outperform three widely-used baselines: Sample Average Approximation (SAA), Empirical Risk Minimization (ERM), and Predict-then-Optimize (PTO), showing consistent advantages across item numbers, service levels, and constraint tightness, as long as the feature dimension is not extremely large; (2) the cost gaps of all methods are decreasing as the resource constraint becomes tighter and these methods perform closely when the resource is very limited; (3) the Empirical Risk Minimization (ERM) method fails to converge and is far away from optimality, even when the sample is relatively large. In a real case study of one of the biggest e-commerce companies in China, our proposed methods outperform all the others, and the inventory costs are significantly reduced.</p>