Features, Labels, and Learned Representations

Machine-learning systems do not encounter a learner, essay, image, or classroom in the same way a person does. They receive representations of observations. Features are the input variables made available to a model, while labels are target values used in many supervised learning tasks. A row in a dataset might contain features such as prior attempts, response time, item history, or encoded text, with a label indicating a later outcome. These choices define what the system can notice and what it is asked to predict.

Some features are selected or constructed directly. Numerical values may be scaled, categories encoded, and text converted into counts or vectors. Feature engineering can make relevant structure easier for a model to use, but it also embeds judgments about what matters. A convenient variable may act as a proxy for socioeconomic status, language background, disability, or institutional opportunity. A timestamp, identifier, or post-outcome record can leak information that would not be available when the system is actually used. High predictive accuracy cannot repair a feature set that makes the intended decision invalid or unfair.

Labels deserve equal scrutiny. They may come from measurements, expert judgments, administrative records, or earlier decisions. A label such as engagement, risk, quality, or mastery is not self-defining. It operationalizes a concept through a particular process. If teachers referred some students for support more often than others, a model trained to reproduce those referrals may learn the referral pattern rather than an independent need. Missing, noisy, delayed, and inconsistently applied labels all shape the learned result.

Modern neural networks often learn internal representations instead of relying only on hand-designed features. During training, successive transformations organize inputs into patterns useful for the objective. An embedding represents an item as a vector so that relationships can be expressed through position and distance. Useful representations can capture regularities that transfer across tasks, but they are not neutral summaries. They reflect the training data, architecture, objective, and feedback. A compact representation may also hide which original details influenced a decision.

Educational evaluation should trace the whole chain from observation to representation to output. Teams should document when each feature becomes available, how labels were created, which groups are represented, and whether performance changes across settings. They should test for leakage, unstable proxies, and distribution shift. Data documentation should preserve these decisions so later reviewers can reconstruct why a representation was considered appropriate. Teachers do not need to inspect every parameter to ask strong questions. What evidence entered the system? What target defined success? What was compressed or omitted? A model learns from the version of reality encoded in its features, labels, and objective, not from the full educational situation those variables are intended to represent.