Every AI model is shaped by the data it learns from, and the kind of data narrows down the kind of model and learning method you can use. This lesson covers the data distinctions the exam names: labeled and unlabeled, structured and unstructured, and the common types: tabular, time-series, image and text.
Labeled and unlabeled data
Labeled data comes with the answer attached. A photo tagged "cat", a past loan marked "repaid", an email marked "spam": each example carries the outcome a model should learn to predict. Unlabeled data has no such answer. It's just the photos, loans or emails.
Labels are valuable but often expensive, because someone usually has to add them by hand, sometimes an expert such as a doctor or lawyer. That's why many real datasets are mostly unlabeled with a small labeled portion. Labeling services exist to help people do this at scale; AWS's, Amazon SageMaker Ground Truth, is now closed to new customers. Whether data is labeled decides the learning type: labeled data enables supervised learning, and unlabeled data points to unsupervised methods.
Structured and unstructured data
Structured data fits neatly into rows and columns with a fixed format, like a customer table with age, region and plan columns. It's easy to search and to feed into traditional ML models.
Unstructured data has no predefined format: images, audio, video, emails, contracts and social media posts. Most of the world's data is unstructured. Getting value from it usually needs deep learning, or a preparation step that pulls structure out of it, such as extracting the text from a scanned form.
Common data types
- Tabular: rows and columns, like a spreadsheet or database table. It's structured, and the classic input for predicting churn, credit risk or prices.
- Time-series: measurements recorded in time order, such as hourly energy use, daily sales or sensor readings each second. The order matters, and it's the natural input for forecasting.
- Image: photos, scans, medical images and video frames. Unstructured, and used by computer vision.
- Text: documents, reviews, chat logs and emails. Unstructured, and used by natural language processing and large language models.
One dataset can mix types. A retailer's records might pair a tabular order history with product photos and customer reviews.
The type of data also points to the type of tool. Tabular data often suits a traditional ML model. Time-series data suits a forecasting model, which pays attention to trends and seasons. Images, audio and free text usually need deep learning or a pre-trained foundation model. So when a scenario describes its data, it's often quietly telling you which answer to pick.
On the exam
Look for the label hiding in plain sight. A spreadsheet with a column such as "cancelled: yes/no" or "defaulted: yes/no" is labeled data, because that column is the answer the model learns to predict. Nobody needs to have tagged it separately. Also watch for timestamps. Data is only "time-series" when the order of readings over time is the point, not just because each row has a date.