What does the local outlier factor capture that a single global density cut-off misses?

## The problem with one global threshold Suppose you score every row by how far it is from its neighbours, then flag everything above a fixed distance. This works only if "far" means the same thing across the whole space. Real data rarely obliges: one region may be a tight, high-volume cluster where neighbours sit within a hair's breadth, and another a legitimately sparse region where ordinary points are far apart. A single cut-off either drowns in false alarms from the sparse region or misses everything unusual inside the tight one. Local outlier factor (Breunig et al., 2000) replaces the absolute measurement with a **relative** one: not "how sparse is this point", but "how sparse is this point *compared with the points it is nearest to*". ## The construction, one layer at a time Fix a neighbourhood size `k`. **k-distance.** For point `p`, `k-distance(p)` is the distance to its k-th nearest neighbour, and `N(p)` is the set of neighbours within that distance. **Reachability distance.** `reach-dist(p, o) = max(k-distance(o), d(p, o))`. Rather than the raw distance from `p` to a neighbour `o`, use at least `o`'s own k-distance. This smoothing stops a point that happens to sit right on top of a dense neighbour from reporting an implausibly tiny distance, which would make the density estimate jumpy. **Local reachability density.** `lrd(p) = 1 / (average reach-dist(p, o) over o in N(p))`. High when the neighbours are close, low when they are far - an inverse-distance density estimate. **The factor itself.** `LOF(p) = average over o in N(p) of ( lrd(o) / lrd(p) )`. Read it as a ratio of densities: - `LOF ~ 1` - the point is about as dense as its neighbours; unremarkable. - `LOF >> 1` - the neighbours live in denser surroundings than the point does; the point is a local outlier. - `LOF < 1` - the point is in a denser spot than its neighbours, which is not an anomaly signal. Because the numerator and the denominator are both densities measured in the same region, the units of density cancel. That is exactly why a cluster's absolute tightness stops mattering. ## A concrete case Hospital billing rows, described by charge amount, length of stay, procedure count and department. A row bills three procedures on a one-day stay. Across the whole hospital thousands of rows look like that, so any global density rule finds it in a crowded part of the space and calls it normal. But among the rows most similar to it - same department, same case mix - that pattern almost never appears: its own nearest neighbours all sit in far denser pockets than it does. The density ratio comes out well above 1 and the row surfaces for review. This is the class of anomaly worth naming when someone asks why a global rule is not enough. ## Choosing k `k` is the one real knob and it is not cosmetic. - **Too small** - the density estimate is dominated by one or two neighbours, so noise produces large factors and the ranking is unstable. - **Too large** - the neighbourhood grows past the local structure you were trying to exploit, the comparison drifts back toward a global one, and small genuine anomaly clusters get absorbed into their own neighbourhoods and hidden. - A practical habit is to pick `k` at least as large as the smallest group of points you would still consider a legitimate cluster rather than an anomaly, and to check that the top of the ranking is stable across a range of `k` rather than trusting one value. ## Costs and cautions - **Cost.** Computing neighbourhoods for every point is quadratic in the number of rows without a spatial index, and spatial indexes themselves degrade as the number of features grows. LOF is a natural fit for tens of thousands of rows, an awkward one for tens of millions. - **It depends entirely on the distance metric.** Features on different scales, one-hot columns and mixed types all distort the neighbourhoods, so the representation has to be built deliberately before LOF means anything. - **High dimensionality.** As features multiply, distances between all pairs of points concentrate toward a similar value, the contrast between "near" and "far" collapses, and every density ratio drifts toward 1. Reducing or selecting features first is often the difference between a useful ranking and noise. - **Scoring new data.** LOF is defined with respect to a reference set of points. Scoring a fresh row means finding its neighbours in that stored reference set, so you carry the data around, and the reference set must be refreshed as the population drifts. - **No natural cut-off.** Values above 1 are suspicious, but there is no threshold that means the same thing on every dataset. Treat the factor as a ranking to be cut by review capacity, not as a calibrated test statistic.