Supervised Learning

The model first produces predictions from its current parameters, then compares those predictions with the known labels. A loss value summarizes the prediction error, and the algorithm uses that signal to adjust its parameters. Repeating this process during training strengthens the learned relationship between inputs and outcomes, improving predictions when the model encounters new data.

Training examples should reflect the kinds of observations the model will encounter later. In behavior research, this may require coverage of relevant actions, behavioral states, environmental conditions, physiological measurements, or digital activity. If the examples do not represent the target data, the learned relationship may perform poorly when applied beyond the training examples.

The labels determine the type of behavioral question the model addresses. Labels representing distinct actions, states, or behavioral categories support classification, whereas labels describing an outcome support prediction of that outcome. In either case, the model learns from paired inputs and labels, allowing researchers to examine behavioral variation through predictions on unseen observations.

Independent data provide a separate basis for judging whether the model learned a useful relationship rather than merely fitting the training examples. Researchers apply the trained model to observations not used during training and compare its predictions with their known labels. This evaluation is especially important when the model will be used on new behavioral data.

A behavior study first identifies input measurements and the labels that represent the actions, states, categories, or outcomes of interest. The labeled examples are then used to train the algorithm, which calculates errors and adjusts its parameters. Finally, researchers apply the trained model to unseen observations and evaluate its performance using independent data.

Possible inputs include direct observations of behavior, physiological measurements, environmental conditions, and digital activity. These sources can be paired with labels describing actions, states, or behavioral categories. Combining such measurements with clearly defined labels allows researchers to examine relationships between recorded conditions and behavioral outcomes within the scope of the available data.

This approach is useful when researchers have examples with known behavioral labels and want to classify observations, predict outcomes, or analyze factors associated with behavioral variation. It can organize information from observations, physiology, environments, or digital activity. Its value depends on representative training data and evaluation on independent observations, rather than training performance alone.