What is noise in data?
Noise is the part of a dataset that varies without meaning. It includes measurement error from sensors, typos and inconsistent entry by people, duplicated records, timestamps in the wrong time zone, and labels assigned differently by different reviewers. Signal is the part of the data that relates to the outcome a model is meant to predict.
Noise sets a ceiling on model performance. If the recorded outcome itself is wrong one time in ten, no model can score better than that on those records. This irreducible error is the third term in the bias-variance tradeoff.
Why noise matters
A model cannot tell noise from signal on its own. Given enough flexibility it will fit the noise, which is overfitting, and then fail on new data where the noise differs. A food plant that feeds raw temperature readings from a line sensor into a quality model may find the model reacting to a sensor that drifts every time a loading door opens. Smoothing the readings, or adding the door state as a feature, removes that noise or turns it into signal.
Label noise is the most expensive kind because it corrupts the ground truth. Two claims adjusters who disagree about what counts as a “complex” claim produce training labels that contradict each other. Data annotation guidelines and a check of agreement between reviewers catch this before training. Data cleansing handles the rest, and anomaly detection can flag records too strange to be real.