Tabular data is regarded as structured data, while other data types such as images, audio, video, text are regarded as unstructured data.

I am confused that, taking images as an example, they are just stored in computers as matrices or high-dimensional tensors, which obviously belong to a certain type of structure, so why are they still called unstructured data?

And even more generally speaking, digital data of any data type should have a certain type of structure, at least from a storage perspective in computers, so what essentially distinguishes between unstructured data and tabular data (structured data)?

And there is also a related term in machine learning called structured learning, which as far as I know refers to learning a model whose output is of dimension higher than 1, like output an image instead of a vector.

But just as mentioned, images are typical unstructured data, so it seems to implicate that the meaning of structured in structured learning conflicts with its meaning in structured data, which confuses me a lot.


1 Answer 1


Tell me the schema for a relation. Maybe students.csv has column attributes of name, gender, height, weight, birthdate, gpa. For each one we can predict what sensible values would be.

For any given row of the relation, I can tell you what the i-th attribute means -- the zeroth is name, and so on. In that way it is "structured".

Now send me puppy1.jpg. What does the zero-th pixel mean? Moving past the upper left corner, what do the next couple of pixels mean? Do they indicate color of background? Color of puppy's fur? Color of mother's fur which it's right next to? Does it indicate ambient lighting level, or time-of-day for a sunlit scene?

Repeat with puppy2.jpg. Does the zero-th pixel have the same meaning?

In this way the pixels are "unstructured" data. They convey some information, but it takes significant effort to understand the context and to discover each pixel's meaning.

  • 3
    $\begingroup$ For an image, single pixel value represents the light intensity at the corresponding position, which is just the physical meaning, and only a group of pixels in a local region can be interpreted as something due to the 2-d structure of an image. But somehow I got what you mean, the concept of structure in structured data refers to the semantic structure rather than the storage structure, right? $\endgroup$
    – XMB-7
    Nov 29, 2023 at 3:19
  • 2
    $\begingroup$ Right. Knowing that the 98th bit of a compressed .CSV influences a student's age might be entertaining, but it's not structured. Decompressing the file and parsing out columns is what lends structure to those stored bits. // Consider a desk with pigeon holes above it, in which we store accounts receivable and accounts payable invoices and such. Several photographs of the desk could be structured data, if the (0, 0) pixel always corresponds to the top left pigeon hole, and we care about their {empty, full} statuses. $\endgroup$
    – J_H
    Nov 29, 2023 at 4:37
  • 1
    $\begingroup$ @XMB-7 you got it right! It's about semantics. $\endgroup$
    – justhalf
    Nov 29, 2023 at 8:52

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.