Tabular vs Unstructured Data

You want to predict loan defaults. Your data is a neat table: one row per borrower, columns for age, income, city.

Your colleague wants to detect spam. Her data is a pile of messy emails.

Your friend wants to find tumors. His data is thousands of X-ray images.

All three are ML problems. But

Not really. Data has families — and the family you are dealing with quietly decides your whole toolchain.

Tabular (structured) data is

Spreadsheets, database tables, and CSV files are the canonical examples: one row per entity, one column per attribute. The schema is fixed, so you always know the meaning of column 7.

Unstructured data is

A single email or a chest X-ray does not fit neatly into columns, so the "features" (the input numbers from topic 28) must be learned or extracted rather than handed to the model directly.

The analogy to carry through:

• Tabular data is a filing cabinet: every drawer labeled, every folder in its place. You can ask a precise question ("all rows where age > 40") and get an instant answer.
• Unstructured data is a shoebox of photos and letters: full of information, but nothing is labeled — you must actually look inside to know what you have.

import pandas as pd
df = pd.DataFrame({
    "age":      [34, 41, 29],
    "income":   [52000, 78000, 44000],
    "city":     ["NYC", "SF", "NYC"],
    "default":  [0, 0, 1],
})
# rows = borrowers, columns = attributes, target = "default"

Question: "will this borrower default?"

Tabular version — three columns of facts:

code
Copy
age  income  city    default
34   52000   NYC     0
41   78000   SF      0
29   44000   NYC     1

Unstructured version — a paragraph of free text:

code
Copy
"I have changed jobs three times this year and
 had to use my emergency credit line twice..."

The table is already numeric-friendly: feed it to a model. The paragraph is not: you must tokenize words, learn meaning, then map to a prediction. Same goal, wildly different plumbing.

Tabular data is the default shape of business and forecasting data — transactions, sensor readings, customer records. It is the domain where classical models still dominate because they handle mixed numeric/categorical columns with little tuning:

• logistic regression
• random forests
• gradient boosting such as XGBoost and LightGBM

Unstructured data powers modern deep learning:

• Text — documents, reviews, chat logs; tokenized into sequences.
• Images — pixels in a grid (H, W, channels); the closest unstructured thing to a numeric array.
• Audio — time-series waveforms, often converted to spectral features (spectrograms) and treated image-like.
• Video / logs — high-dimensional and temporally structured.

Models used per family:

• Images → convolutional networks (CNNs) or vision transformers.
• Text and speech → transformer language models.
• Long-range dependencies in sequences → recurrent networks (RNNs).

The distinction is not cosmetic — it changes the toolchain at every layer:

code
Copy
tabular       CSV/DB  ->  pandas/SQL/Parquet  ->  scikit-learn (XGBoost)
unstructured  files   ->  blob store / lake    ->  PyTorch / TensorFlow
                        (tokens, pixels,       (CNN / Transformer)
                         waveforms)

• Storage: tabular work lives in pandas, SQL and Parquet; unstructured work lives in object storage or a data lake.
• Models: tabular feeds scikit-learn-style estimators; unstructured feeds PyTorch/TensorFlow deep networks.
• Preprocessing: tabular needs imputation and encoding; unstructured needs tokenization, resizing and augmentation.
• Hardware: classical tabular models often train on CPUs; deep unstructured models realistically want GPUs.

Tabular:     age=34, income=52000, city=NYC   -> LogisticRegression
Unstructured: "The movie was a total waste..." -> BERT -> sentiment class

The line is blurring.

• Tabular deep-learning models and embedding pipelines now compete with gradient boosting on medium-sized tables.
• Feature stores serve derived features from unstructured sources back into tabular models — e.g., an image classifier outputs a venue-style score that becomes just another column.

Carrying the analogy forward: you now run the shoebox of photos through a deep model first, and it hands you a few tidy numbers you can file in the cabinet.

Still, most production forecasting, risk and pricing systems remain fundamentally tabular — which is why this course builds tabular fluency first. Knowing which family your data belongs to saves you from the two classic mistakes: forcing a neural net onto a small table, or forcing a gradient-boosted tree onto raw pixels.