#tabular-data

Question 1: Single choice

A bank has about 5,000 rows of tabular customer data (age, income, account balance, …) and wants to predict loan default. Which plan is best supported by benchmark evidence on this kind of data?

Fit a strong conventional model, such as gradient-boosted trees, as a baseline any deep network must beat1
Train a deep network from scratch at once, since deep models outperform conventional ones on every kind of data2
Fine-tune a network pretrained on photos, because pretrained features carry over to any type of data3
Skip the baseline, because comparing with simpler models matters only for image data4

Timed test