What is an Outlier?
Outliers, told as a story about spotting the one piece of produce that stands out from a group of otherwise similar ones.
Among similarly sized apples, one stands out — much bigger.Even at a glance, the big one is clearly different from the rest.
Ignore an outlier and the average skews; discard it blindly and an important signal may be lost.An outlier might be a mistake — or it might be a genuinely important signal.
An outlier is a data point that's statistically, clearly unlike the rest of the group.'Different' isn't a feeling — it's measured by distance or spread.
There are different ways to find an outlier — but they all come down to agreeing on a cutoff line.Where that line is drawn decides whether the very same item counts as an outlier or not.
Not every outlier is meant to be thrown away.An item tampered with by an adversarial attack can also look statistically like an outlier.
An outlier = a data point that, measured by distance or standard deviation, is clearly unlike the rest of the group.
An outlier is a value clearly unlike the rest of the data, found using statistical criteria like distance from a cluster's center, Z-score (standard deviations from the mean), or falling far outside the IQR (interquartile range). It can be the trace of a sensor error or a data-entry mistake — or it can be a rare but genuinely important signal, like a fraudulent transaction, a rare disease, or the trace of an attack. The rule of thumb is to investigate the cause before deleting it, not the other way around. The 'noise' that DBSCAN clustering sets aside is itself one face of an outlier.
When grown-ups say it
- Outlier
- The piece of produce that stands out from the crowd. Found by measuring distance or standard deviation.
- Z-score
- The number of standard deviations a value sits from the mean. For roughly bell-shaped data, an absolute value past 3 is commonly treated as an outlier.
- IQR (interquartile range)
- The range where the middle 50% of the data sits. Falling more than 1.5 times this range's width beyond it commonly marks an outlier.
- Noise
- What DBSCAN leaves aside, belonging to no cluster. How an outlier shows up inside clustering. → Grouping by Likeness, With No Name Tags
- Anomaly detection
- Making the hunt for outliers the whole point of the task. Used for things like fraud detection and predicting equipment failure.
- Data cleaning
- Weeding out outliers confirmed to be errors. One step in managing the quality of the seed — the training data. → A Sorter Trained on Rotten Seed
- Adversarial attack
- A tampering that can also look statistically like an outlier. One defense against it is exactly this kind of detection. → The Fake Dot Pattern Snuck Onto the Produce
- Confusion matrix
- An outlier detector's own performance is also checked with this table. What it missed can be read the same way, in four boxes. → The Four-Box Stamp Ledger