Tabular Data
1. Purpose & Motivation
1.1 Why tabular data matters
- Most enterprise and scientific data is stored in tables (databases, spreadsheets, CSVs)
- Tabular data is the most common data format in real-world ML applications
- Understanding its structure is essential for choosing appropriate models and preprocessing
1.2 The core problem it addresses
Organize observations as rows and attributes as columns to enable systematic analysis, modeling, and storage of structured information.
2. Definition & Core Idea
2.1 Formal definition
Tabular data is data organized in a two-dimensional table where each row represents an observation (sample, record) and each column represents a feature (attribute, variable). Each cell contains a single value.
2.2 Intuitive mental model
"A spreadsheet where each row is a patient and each column is a measurement (age, blood pressure, diagnosis) -- the most natural way to organize structured information."
3. Structure
3.1 Anatomy of a table
| Component | Role | Example |
|---|---|---|
| Row | One observation / sample | A single customer |
| Column | One feature / attribute | Customer age |
| Cell | Single value at a row-column intersection | 34 |
| Index | Unique identifier per row | Customer ID |
| Header | Column names describing each feature | "age", "income" |
3.2 Feature types in tabular data
| Type | Description | Examples |
|---|---|---|
| Numerical | Continuous or discrete numbers | Price, count, temperature |
| Categorical | Finite set of discrete labels | Color, country, product type |
| Ordinal | Categorical with meaningful order | Education level, rating (1-5) |
| Boolean | Binary true/false | Is_active, has_subscription |
| DateTime | Temporal values | Purchase date, signup timestamp |
| Text | Free-form strings | Product description, comments |
4. Common Preprocessing Steps
4.1 Handling missing values
- Imputation (mean, median, mode, or model-based)
- Dropping rows or columns with excessive missingness
- Indicator columns flagging that a value was missing
4.2 Encoding and scaling
- One-hot or target encoding for categorical features
- Standardization or normalization for numerical features
- Log transforms for skewed distributions
4.3 Data quality checks
- Duplicate detection and removal
- Outlier identification
- Consistency validation across columns
5. Models That Excel on Tabular Data
| Model family | Why it works well |
|---|---|
| Gradient Boosting | State-of-the-art; handles mixed types natively |
| Random Forest | Robust, minimal preprocessing required |
| Linear models | Fast, interpretable baseline |
| Deep learning | Competitive but often requires more tuning |
Key insight: Tree-based ensemble methods (XGBoost, LightGBM, CatBoost) consistently dominate benchmarks on tabular data.
6. Tabular Data vs Other Data Types
| Aspect | Tabular | Images / Text / Audio |
|---|---|---|
| Structure | Rows and columns | Grid, sequence, waveform |
| Features | Heterogeneous types | Homogeneous (pixels, tokens) |
| Best models | Tree ensembles | Deep neural networks |
| Preprocessing | Feature engineering | Learned representations |
7. One-Sentence Summary
Tabular data organizes observations as rows and features as columns, forming the most common data format in real-world machine learning where tree-based models dominate.