Outlier

An outlier is a data point that does not fit the pattern formed by the rest of the data. A useful way to think about it is: if most observations form a crowd, an outlier is the one standing noticeably apart. In unsupervised learning, that “apartness” is judged without labels, so the model has to infer what looks normal from the data itself.

An outlier is not just a large or unusual value in one column. It is an observation whose distance, density, or probability under the learned data pattern is markedly different from typical points. Different methods define this differently:

• Distance-based methods flag points far from their neighbors.
• Density-based methods, such as Local Outlier Factor (LOF), look for points sitting in sparse regions compared with nearby areas.
• Model-based methods, such as Isolation Forest or One-Class SVM, learn what normal data looks like and score points that break that pattern.

Outliers matter because they can be the whole point of the analysis or a serious source of distortion. In fraud detection, a strange transaction can signal theft. In manufacturing, an unusual sensor reading can indicate a failing machine. In customer data, an outlier might be a VIP buyer—or a data entry error. If ignored, outliers can pull cluster centers in the wrong direction, stretch scales, and make dimensionality reduction methods like PCA focus on rare noise instead of useful structure.

In practice, outliers are usually identified through an anomaly score rather than a simple yes/no rule, and then compared against a threshold. In Python, this shows up in tools like sklearn.ensemble.IsolationForest and sklearn.neighbors.LocalOutlierFactor. The key idea is simple: an outlier is a point whose relationship to the rest of the dataset is so different that it deserves special attention, either because it reveals something important or because it can mislead the rest of the analysis.