AI and machine learning for tabular data: a practical guide
Artificial intelligence is the broad goal of building machines that do tasks that need intelligence. Machine learning is the main way to get there today: models that learn patterns from data instead of following hand-written rules. Deep learning is a part of machine learning that uses neural networks with many layers. Tabular data, the rows and columns in databases and spreadsheets, is where most business data lives, and it has its own best practices.
The learning paradigms, and what they mean for tables
- Supervised learning learns from labeled rows. It is the standard setting for tabular prediction.
- Unsupervised learning finds structure without labels, for example clusters or anomalies.
- Self-supervised learning creates its own labels from the data, for example by masking a cell and predicting it. It is how tabular foundation models are pre-trained.
- Transfer learning reuses what a model learned on one task for another. A pre-trained model fine-tuned on your table is an example.
- Few-shot learning learns from a handful of labeled rows, which is where in-context prediction helps.
- Federated learning trains across parties without sharing raw data.
- Reinforcement learning learns from rewards over time and is used for decisions, not for plain prediction.
Trees: the baseline you should beat
XGBoost, CatBoost, and the random forest are strong, fast, and well understood on tabular data. They need little preprocessing, handle mixed types, and work well on small and medium datasets. Any new method, deep or not, should be compared against a tuned tree ensemble on the same split.
Deep models for tables
Deep models such as FT-Transformer and SAINT apply attention to tabular features. They can match trees on some datasets but are not reliably better, and they need more tuning. Their value is that they learn representations that can be shared across tasks, which led to tabular foundation models.
Tabular foundation models
A tabular foundation model is pre-trained once on many tables, so a new table needs no training run. A large tabular foundation model adds scale and several capabilities in one checkpoint: prediction, imputation, anomaly detection, and matching.
How to choose
- One stable dataset, time to tune, strict latency: start with gradient-boosted trees.
- Many different tables, little time, unnamed or inconsistent columns: try a tabular foundation model.
- Missing values are the main problem: compare imputation on your own data with a masked-cell test.
- You need several outputs from one table: a single large model is simpler than several pipelines.
In every case, hold out data the models have never seen, use the same split for every method, and report how each method was tuned. A score is only as good as its report explains why.
Frequently asked questions
What is the difference between AI and machine learning?
AI is the goal. Machine learning is the main method: models that learn from data.
Is deep learning better than XGBoost for tabular data?
Not reliably. Tree ensembles are strong baselines. Deep and foundation models help most when you have many tables or little time to tune.
When should I try a tabular foundation model?
When you work across many tables, have unnamed columns, need imputation or anomaly detection too, or cannot afford per-dataset tuning.