The Local Outlier Factor (LOF) filter was introduced as a method of determining the degree to which an object is an outlier. This filter is an implementation of the method described in [Breunig2000].

The filter creates three new dimensions, KDistance, LocalReachabilityDistance and LocalOutlierFactor, all of which are double-precision floating values. The KDistance dimension records the Euclidean distance between a point and it’s k-th nearest neighbor (the number of k neighbors is set with the minpts option). The LocalReachabilityDistance is the inverse of the mean of all reachability distances for a neighborhood of points. This reachability distance is defined as the max of the Euclidean distance to a neighboring point and that neighbor’s own previously computed KDistance. Finally, each point has a LocalOutlierFactor which is the mean of all LocalReachabilityDistance values for the neighborhood. In each case, the neighborhood is the set of k nearest neighbors.

In practice, setting the minpts parameter appropriately and subsequently filtering outliers based on the computed LocalOutlierFactor can be difficult. The authors present some work on establishing upper and lower bounds on LOF values, and provide some guidelines on selecting minpts values, which users of this filter should find instructive.


To inspect the newly created, non-standard dimensions, be sure to write to an output format that can support arbitrary dimensions, such as BPF.

Default Embedded Stage

This stage is enabled by default


The sample pipeline below computes the LOF with a neighborhood of 20 neighbors, followed by a range filter to crop out points whose LocalOutlierFactor exceeds 1.2 before writing the output.




The number of k nearest neighbors. [Default: 10]


An expression that limits points passed to a filter. Points that don’t pass the expression skip the stage but are available to subsequent stages in a pipeline. [Default: no filtering]


A strategy for merging points skipped by a ‘where’ option when running in standard mode. If true, the skipped points are added to the first point view returned by the skipped filter. If false, skipped points are placed in their own point view. If auto, skipped points are merged into the returned point view provided that only one point view is returned and it has the same point count as it did when the filter was run. [Default: auto]