Data is any information that can be collected and used to train a machine learning model. Not all data looks the same — some is organized in tables, some is text or images, and some is a mix.
The type of data you have determines what preprocessing steps are needed, which model architectures work best, and how you measure performance.
Images — Require resizing, normalization, and special neural networks
Text — Requires tokenization and word embeddings
Tables — Require handling missing values and encoding categories
Audio — Requires conversion to spectrograms
Structured data is organized in rows and columns, like a table. It is easy to store, query, and use with most machine learning algorithms.
Features — House Size, Bedrooms, Location
Label — Price
Each row is one house. Each column is one feature. This clear structure makes it easy for algorithms to learn.
Unstructured data does not follow a fixed format. It requires special processing before most models can use it.
Sentiment analysis — Positive or negative tone
Topic classification — What the text is about
Language translation — Converting between languages
Question answering — Responding to queries
Object detection — Finding objects in an image
Facial recognition — Identifying people
Medical image analysis — Detecting diseases in scans
Autonomous vehicles — Understanding road scenes
Speech recognition — Converting speech to text
Emotion detection — Identifying emotional tone
Speaker identification — Recognizing who is speaking
Music genre classification — Labeling music styles
Action recognition — Identifying human activities
Surveillance analysis — Detecting unusual events
Content moderation — Filtering inappropriate content
Semi-structured data has some organization but does not fit neatly into rows and columns. It requires parsing before most models can use it.
name: "Alice"
age: 28
city: "Toronto"
This is structured within the object but different records may have different fields.
| Type | Format | Examples | Common Use |
|---|---|---|---|
| Structured | Rows and columns | Spreadsheets, databases | Classification, regression |
| Unstructured | No fixed format | Text, images, audio | NLP, computer vision |
| Semi-Structured | Partial format | JSON, XML, email | Web APIs, document parsing |
Numerical data consists of numbers. There are two types.
Can take any value in a range
Only specific whole number values
Categorical data consists of labels or groups. Most models require categorical data to be converted into numbers before training.
Blood type — A, B, AB, O
Color — Red, Blue, Green
Country — Canada, USA, Mexico
Product category — Electronics, Clothing, Food
Time-series data is recorded in sequence over time. Each data point is associated with a specific timestamp.
2026-06-01: 22°C
2026-06-02: 24°C
2026-06-03: 19°C
2026-06-04: 21°C
Time-series data helps models predict future values based on past patterns.
CSV — Comma-separated values, common for structured data
JSON — Flexible format for semi-structured data
XML — Used in documents and configurations
JPEG/PNG — Image formats
MP3/WAV — Audio formats
SQL databases — For structured relational data
| Data Type | Common Model Types |
|---|---|
| Structured (tabular) | Decision Trees, Random Forests, Linear Regression |
| Text | Transformers, RNN, LSTM |
| Images | Convolutional Neural Networks (CNN) |
| Audio | CNN on spectrograms, RNN |
| Time-series | LSTM, ARIMA, Transformers |
Structured — User purchase history tables
Text — Product reviews
Images — Product photos
Semi-structured — JSON user profiles
Structured — Patient records and lab results
Images — X-rays and CT scans
Text — Doctor notes and reports
Audio — Heart and lung sounds
Structured — Missing values, inconsistent formats, outliers
Text — Ambiguity, sarcasm, multiple languages
Images — Lighting variation, size differences, partial occlusion
Audio — Background noise, speaker variation, accents
Video — High computational cost, tracking across frames
Ask the AI if you need help understanding structured, unstructured, or semi-structured data, numerical versus categorical data, or how to choose the right preprocessing steps for different data types.