Types of Data

The hospital intake form

A patient arrives at a hospital. Before any treatment begins, a nurse fills out an intake form. Every field is a measurement — but not every measurement is the same kind of thing.

The form collects: age (47 years), body temperature (37.8°C), blood type (AB), pain level (7 out of 10), ward assignment (cardiology), smoker status (yes), and number of previous admissions (2). Seven fields. Four fundamentally different types of data.

An ML system that treats these seven fields identically will make serious errors. The algorithm does not know that 7 on a pain scale is not seven times worse than 1, or that blood type AB is not numerically between A and O. It only sees the numbers it is given — and if those numbers carry false implications, the algorithm will act on them.

import pandas as pd

df = pd.read_csv('hospital_intake.csv')
print(df.dtypes)
# age              float64    ← continuous
# temperature      float64    ← continuous
# blood_type       object     ← categorical (stored as text)
# pain_level       int64      ← ordinal (looks like integer!)
# smoker           bool       ← binary
# prior_admissions int64      ← count

Python’s dtype reports how data is stored, not what kind of data it is statistically. A pain score stored as int64 is not truly continuous just because Python treats it as a number — and treating it as one leads to the encoding bugs covered at the end of this chapter.

Continuous data

Age and temperature are continuous measurements. A continuous variable can take any value within a range, including all the decimal fractions in between. A patient can be 47.3 years old. A temperature can be 37.83°C. The number line for these measurements is smooth and unbroken — there are no gaps between valid values.

Because the gaps are equal and meaningful, continuous data supports all arithmetic operations. You can meaningfully average two temperatures, subtract one age from another, or compute the distance between two patients in feature space. Algorithms that rely on distance — KNN, SVMs with RBF kernels, neural networks — work directly on continuous data without any transformation.

Categorical data

Blood type and ward assignment are categorical measurements. A patient’s blood type is one of four labels: A, B, AB, or O. These labels have no natural order. O is not “more” than A. AB is not between A and O. Neurology is not halfway between cardiology and oncology. The only relationship between categories is same versus different.

For categorical features with hundreds or thousands of unique values — zip codes, product IDs, user accounts — one-hot encoding creates extremely wide, sparse vectors. Modern ML systems often learn compact dense embeddings instead, training a short vector for each category value alongside the main model. This approach is covered in the Categorical Encoding chapter of the Features & Representations topic.

Ordinal data

Pain level is ordinal. The values run from 1 to 10 and they have a natural order — a patient reporting 8 is in more pain than one reporting 4. But the gaps between values are not equal or meaningful. The jump from level 3 to level 4 may not represent the same increase in pain as the jump from level 8 to level 9. The numbers encode rank, not magnitude.

Ordinal data sits in an awkward position between categorical and continuous. Tree-based algorithms (decision trees, random forests, gradient boosting) handle ordinal data naturally — they split at thresholds without assuming equal gaps. Distance-based algorithms treat ordinal values as continuous unless told otherwise, which can introduce errors.

Binary data

Smoker status is binary: yes or no, with no other options. Binary data is a special case of categorical data with exactly two values. It gets its own category because it can be encoded as a single 0/1 column without implying any false mathematical relationship. 1 is not “greater” than 0 in a meaningful sense — 1 simply means “smoker”.

Count data

The number of previous admissions is a count variable: a non-negative integer (0, 1, 2, 3…). Count data resembles continuous data but has important constraints — it cannot be negative, it has no decimal values, and its distribution is often skewed right. Poisson regression and related methods treat count data as its own type.

Why this matters for ML

Every algorithm that uses distance — KNN, SVMs, neural networks, k-means clustering — assumes that numbers close together mean similar things. If you encode blood type as O=4 and AB=3, the algorithm will treat those two types as more similar to each other than A=1 and O=4, because |3−4|=1 while |1−4|=3. This relationship is not biological reality — it is an artefact of the encoding choice.

The data type determines the encoding. The encoding determines what the algorithm sees. Getting this wrong is invisible — no error is raised, the model trains, predictions are produced, and the silent bug propagates into every downstream decision.

The shape of data types

Each data type has a visual signature when sampled at scale. Continuous measurements form smooth, bell-shaped or skewed curves — age clusters between 30 and 70, temperature hovers near 37°C, with no empty gaps in the distribution. Categorical fields form discrete, isolated bars — four bars for blood type with no values in between. Binary fields produce exactly two bars. Ordinal fields look like discrete bars with a meaningful left-to-right ordering.