Identifying the types of data used in machine learning, model training and ethical AI deployment.
At the Canadian AI Advancement Institute (CAIAI), our mission is to help Canadians learn, adapt, and succeed with AI through accessible AI education and training. As organizations increasingly use artificial intelligence and machine learning to automate decisions, personalize services, and derive insights from their data, understanding how data functions within these systems is essential.
In this blog, we explore the types of data used in machine learning, including structured data, unstructured data, labelled data and unlabelled data, alongside best practices for data selection and key AI ethics considerations that help ensure responsible, trustworthy AI systems.
This blog explains the differences between structured data and unstructured data, showing how rows-and-columns information, such as transaction logs and customer databases, contrasts with freeform content like social media posts, audio, images, and video. It then explores how labelled data and unlabelled data support different learning approaches, from supervised tasks such as spam detection to unsupervised pattern discovery in customer behaviour.
Building on these foundations, it outlines practical principles for effective data selection, including relevance, accuracy, diversity, completeness, and consistency, as well as the role of preprocessing in cleaning and preparing datasets for training. Lastly, it explores several AI ethics considerations, discussing how poor data practices can lead to harmful or biased outcomes.
At the core of any machine learning system is data, and understanding the different data types will help you use the right techniques and avoid costly mistakes. Broadly, we can think about four key categories: structured data, unstructured data, labelled data, and unlabelled data.
Structured data is highly organized information that fits neatly into tables with rows and columns. Examples include transaction logs, sensor readings, customer databases and charted statistics. Because structured data is standardized, it is easy for machine learning algorithms to query, filter, and analyze. For instance, a financial institution might use structured data such as credit scores, income levels, and payment history to build a risk model to assess loan eligibility or detect unusual account activity.
By contrast, unstructured data does not follow a fixed format. It includes social media posts, emails, audio recordings, images, and videos. A single dataset might combine text comments with attached photos or video clips, each containing useful data but lacking a consistent structure. To derive value from this content, machine learning models must first interpret the raw data. AI tools, including consumer-facing multimodal models such as ChatGPT, Gemini, and Perplexity, and/or image recognition tools, can interpret this unstructured data and uncover patterns that would be impossible to see at a glance.
Building on this, we can further distinguish between labelled data and unlabelled data.
Labelled data contains annotations or tags that tell the model what each example represents. In a dataset of animal photos, labels might indicate “dog,” “cat,” or “horse.” In an email dataset, each message might be labelled as “spam” or “not spam.” These labels serve as ground truth, enabling supervised machine learning models to learn a direct mapping from inputs (data) to outputs (labels). This is essential for tasks like classification, prediction, and regression.
Unlabelled data, on the other hand, consists of raw inputs with no such annotations. Examples include social media posts without sentiment tags, raw audio recordings, or unannotated video footage. In these cases, the model is not told what to look for. Instead, it must explore the data to find patterns on its own. This is the basis of unsupervised learning, where algorithms might cluster customers by browsing behaviour or group similar documents together based on their content.
In practice, most real-world machine learning pipelines combine these types. Organizations might use structured, labelled data for highly regulated tasks such as credit scoring, while drawing on large volumes of unstructured, unlabelled data, such as customer feedback surveys, to identify trends.
Once we understand what kinds of data are available, the next step is data selection. Not all data is equally useful, and more data is not always better. Effective data selection means choosing information that supports the specific problem the machine learning model is meant to solve. Ideally, the data used in an AI system should be relevant, accurate, diverse, complete and consistent.
Let’s dive into these data qualities further:
After appropriate data selection, preprocessing prepares the data for machine learning. This step includes tasks such as cleaning, standardizing and augmenting the selected data.
Let’s take a look at the specifics of each of these initiatives:
Without thoughtful preprocessing, even high-quality structured or unstructured data can produce inconsistent or biased models. The model may latch onto noise or negative attributes within the dataset, resulting in unreliable predictions upon deployment.
Ethical considerations must be considered throughout the data selection and machine learning process. From collection and storage to training and deployment, employing ethical AI principles will directly shape the safety, fairness, and trustworthiness of your machine learning systems. Let’s take a closer look at some key ethical considerations specific to the data collection process.
First, privacy is critical. Many datasets may contain sensitive information about individuals, including financial details, health records, and personal identifiers. Organizations must implement safeguards such as anonymization, access controls, and secure storage, and comply with applicable privacy and data protection laws.
Second, consent matters. People and organizations should understand how their data is collected and used, and they should have the opportunity to consent or decline to participate. Clear consent processes help build trust and reduce the risk of misuse.
Third, data quality and oversight are central to ethical AI. Not all data is appropriate for training machine learning models. Harmful, misleading, or manipulated data can lead to systems that reproduce or amplify those issues.
As organizations continue to explore the benefits of AI, developing literacy in machine learning and the data used in AI systems is an essential first step toward responsible and effective adoption.
For individuals and teams seeking machine learning education and training in artificial intelligence, the Canadian AI Advancement Institute offers a range of opportunities, including AI education, an Introduction to Artificial Intelligence course, AI workshops, and AI training tailored to institutional, industry, and government needs.
Whether you are beginning your learning journey or preparing your organization for broader AI adoption, CAIAI offers a range of educational pathways to help you learn, adapt, and succeed with AI.
