Tabular data, 50k rows, 30 features. Neural network?

Almost certainly not. At that scale — 50,000 rows, 30 features — gradient boosting (XGBoost, LightGBM, that family) will usually match or beat a neural network, train in minutes instead of hours, and remain explainable with feature-importance methods. That last point matters more than it sounds: if a decision ever needs to be defended to a regulator or a client, “here’s the SHAP breakdown” is a very different conversation than “here’s what a neural net’s weights encode.”

Deep learning earns its place on unstructured data — images, text, raw sequences — or specifically where learning a representation is the point, not on small structured tables where a well-tuned tree ensemble is both simpler and stronger.