In the domains of machine learning and data science, the term ground truth gets thrown around a lot. As a beginner in the field, I found myself being unable to follow explanations because I was unsure of the term. What exactly is ground truth, and why is it useful? I decided to do a little research, and piece together a definition from a variety of sources.
In a nutshell, ground truth is the observable data that is gathered empirically. When building predictive models, it's the data that the models are compared against in order to determine how accurate they are. Ground truth can - and almost invariably does - contain noise.
When building predictive models, the ground truth is the data that the models are compared against in order to determine how accurate they are.
Ground truth looks different in different application domains of machine learning and data science. Here are some examples of what ground truth looks like for automated driving, financial markets, and speech recognition applications.
- In a dataset for automated driving, the ground truth is the footage that the video camera captured when driving around in an automobile. The dataset includes the class labeling for what a pedestrian is, what another automobile is, what is the road, etc. There is a proposal for using videogame footage as ground truth for automated driving applications.
- In financial markets, the ground truth is the raw market data that we are trying to predict, and that we are using to base our predictions from. For example in the stock market, the ground truth of a particular stock is what the price was, and when.
- In speech recognition, the ground truth is made up of massive amounts of data of human voice recordings, in all sorts of different spaces with different reverberation qualites.