Machine learning professionals use structured prediction in a whole multitude of ways, typically by applying some form of machine learning technique to a particular goal or problem that can benefit from a more ordered starting point for predictive analysis.
A technical definition of structured prediction involves “predicting structured objects rather than scalar discrete or real values.”
Another way to say that would be that instead of simply measuring individual variables in a vacuum, structured predictions work from a model of a particular structure, and use that as a basis for learning and making predictions. (Read How Can AI Help in Personality Prediction?)
The techniques for structured prediction are widely variable – from Bayesian techniques to inductive logic programming, Markov logic networks and structured support vector machines or nearest neighbor algorithms, machine learning professionals have a broad toolset at their disposal to apply to data problems.
What's common in these ideas is the use of some underlying structure that the machine learning work is founded on inherently.
Experts often give the idea of natural language processing, where parts of speech are tagged to represent elements of a text structure – other examples include optical character recognition, where a machine learning program recognizes handwritten words by parsing segments of a given input, or complex image processing, where computers learn to recognize objects based on segmented input, for example, with convolutional neural network comprised of many “layers.”
Experts might talk about linear multiclass classification, linear compatibility functions and other basis techniques for generating structured predictions. In a very general sense, structured predictions build on a different model than the wider field of supervised machine learning — to go back to the example of structured predictions in natural language processing and tagged phonemes or words, we see that the use of the labeling for supervised machine learning is oriented toward the structural model itself — the meaningful text that is supplied, perhaps in test sets and training sets.
Then, when the machine learning program is let loose to do its work, it's founded on the structural model. That, experts say, explains some of how the program understands how to utilize parts of speech like verbs, adverbs, adjectives and nouns, rather than mistaking them for other parts of speech, or not being able to distinguish how they work in a global context. (Read How Structured Is Your Data? Examining Structured, Unstructured and Semi-Structured Data.)
The field of structured prediction remains a key part of machine learning as various types of machine learning and artificial intelligence evolve.