For images, text and audio, deep learning dominates. For tabular data — rows and columns in business databases — the picture is different.
Tree-Based Models Often Win
Gradient-boosted trees such as XGBoost, LightGBM and CatBoost frequently match or beat neural networks on tabular benchmarks. Reasons include:
- Handling of mixed feature types and irregular patterns.
- Robustness to uninformative features.
- Less tuning needed.
- Fast training on modest hardware.
When Deep Learning Helps
- Very large datasets.
- Tabular data combined with text, images or sequences.
- Learning embeddings for high-cardinality categories.
- Transfer learning across related tables.
- Newer tabular foundation models, which can perform well on small datasets.
Practical Advice
- Start with gradient-boosted trees as a strong baseline.
- Try neural approaches if the data is large or multimodal, or if baselines plateau.
- Compare fairly, with equal tuning effort.
Beyond Accuracy
Consider training cost, inference speed, interpretability and maintenance.