Density-based Approaches
Why Density-Based Clustering methods?
Discover clusters of arbitrary shape.
Clusters – Dense regions of objects separated by
regions of low density
DBSCAN – the first density based clustering
OPTICS – density based cluster-ordering
DENCLUE – a general density-based description
of cluster and clustering
Source: https://cse.buffalo.edu/faculty/azhang/cse601/density-based
.ppt
DBSCAN: Density Based Spatial
Clustering of Applications with Noise
Proposed by Ester, Kriegel, Sander, and Xu
(KDD96)
Relies on a density-based notion of cluster: A
cluster is defined as a maximal set of density-
connected points.
Discovers clusters of arbitrary shape in spatial
databases with noise
Density-Based Clustering
Basic Idea:
Clusters are dense regions in the
data space, separated by regions of
lower object density
Why Density-Based Clustering?
Results of a k-medoid
algorithm for k=4
Different density-based approaches exist (see Textbook & Papers)
Here we discuss the ideas underlying the DBSCAN algorithm
Density Based Clustering: Basic Concept
Intuition for the formalization of the basic idea
For any point in a cluster, the local point density
around that point has to exceed some threshold
The set of points from one cluster is spatially
connected
Local point density at a point p defined by two
parameters
– radius for the neighborhood of point p:
N (p) := {q in data set D | dist(p, q) }
MinPts – minimum number of points in the given
neighbourhood N(p)
-Neighborhood
-Neighborhood – Objects within a radius of
from an object.
N ( p ) : {q | d ( p, q ) }
“High density” - ε-Neighborhood of an object
contains at least MinPts of objects.
ε-Neighborhood of p
ε ε ε-Neighborhood of q
qq pp
Density of p is “high” (MinPts = 4)
Density of q is “low” (MinPts = 4)
Core, Border & Outlier
Outlier Given and MinPts,
categorize the objects into
Border three exclusive groups.
A point is a core point if it has more than
Core a specified number of points (MinPts)
within Eps These are points that are at
the interior of a cluster.
A border point has fewer than
= 1unit, MinPts = 5 MinPts within Eps, but is in the
neighborhood of a core point.
A noise point is any point that is not
a core point nor a border point.
Example
M, P, O, and R are core objects since each is
in an Eps neighborhood containing at least 3
points
Minpts = 3
Eps=radius
of the
circles
Density-Reachability
Directly density-reachable
An object q is directly density-reachable from
object p if p is a core object and q is in p’s -
neighborhood.
q is directly density-reachable from p
ε ε p is not directly density- reachable
qq pp from q?
Density-reachability is asymmetric.
MinPts = 4
Density-reachability
Density-Reachable (directly and
indirectly):
A point p is directly density-reachable from p2;
p2 is directly density-reachable from p1;
p1 is directly density-reachable from q;
pp2p1q form a chain.
p p is (indirectly) density-reachable
p2 from q
p1 q is not density- reachable from p?
q
MinPts = 7
Density-Connectivity
Density-reachable is not symmetric
not good enough to describe clusters
Density-Connected
A pair of points p and q are density-connected
if they are commonly density-reachable from a
point o.
Density-connectivity is
symmetric
p q
o
Formal Description of Cluster
Given a data set D, parameter and threshold
MinPts.
A cluster C is a subset of objects satisfying
two criteria:
Connected: ∀ p,q ∈ C: p and q are density-
connected.
Maximal: ∀ p,q: if p ∈ C and q is density-
reachable from p, then q ∈ C. (avoid redundancy)
P is a core object.
Review of Concepts
Is an object o in a cluster Are objects p and q in the
or an outlier? same cluster?
Are p and q density-
Is o a core object?
connected?
Is o density-reachable by Are p and q density-
some core object? reachable by some
object o?
Directly density- Indirectly density-
reachable reachable through a chain
DBSCAN Algorithm
Input: The data set D
Parameter: , MinPts
For each object p in D
if p is a core object and not processed then
C = retrieve all objects density-reachable from p
mark all objects in C as processed
report C as a cluster
else mark p as outlier
end if
End For
DBScan Algorithm
DBSCAN: The Algorithm
Arbitrary select a point p
Retrieve all points density-reachable from p wrt Eps and
MinPts.
If p is a core point, a cluster is formed.
If p is a border point, no points are density-reachable from p
and DBSCAN visits the next point of the database.
Continue the process until all of the points have been
processed.
DBSCAN Algorithm: Example
Parameter
= 2 cm
MinPts = 3
for each o D do
if o is not yet classified then
if o is a core-object then
collect all objects density-reachable from o
and assign them to a new cluster.
else
assign o to NOISE
DBSCAN Algorithm: Example
Parameter
= 2 cm
MinPts = 3
for each o D do
if o is not yet classified then
if o is a core-object then
collect all objects density-reachable from o
and assign them to a new cluster.
else
assign o to NOISE
DBSCAN Algorithm: Example
Parameter
= 2 cm
MinPts = 3
for each o D do
if o is not yet classified then
if o is a core-object then
collect all objects density-reachable from o
and assign them to a new cluster.
else
assign o to NOISE
DBSCAN Algorithm
Eliminate noise points
Perform clustering on the remaining points
18
MinPts = 5
P1
C1 P
P C1
C1
1. Check the - 1. Check the unprocessed
neighborhood of p; objects in C
2. If p has less than MinPts 2. If no core object, return C
neighbors then mark p
3. Otherwise, randomly pick
as outlier and continue
up one core object p1,
with the next object
mark p1 as processed,
3. Otherwise mark p as and put all unprocessed
processed and put all neighbors of p1 in cluster
the neighbors in cluster C
C
C1
C1
C1 C1
C1
Example
Original Points Point types: core,
border and outliers
= 10, MinPts = 4
When DBSCAN Works Well
Original Points Clusters
• Resistant to Noise
• Can handle clusters of different shapes and sizes
When DBSCAN Does NOT Work Well
(MinPts=4, Eps=9.92).
Original Points
• Cannot handle Varying
densities
• sensitive to parameters
(MinPts=4, Eps=9.75)
DBSCAN: Sensitive to Parameters
Determining the Parameters
and MinPts
Cluster: Point density higher than specified by and MinPts
Idea: use the point density of the least dense cluster in the data
set as parameters – but how to determine this?
Heuristic: look at the distances to the k-nearest neighbors
p 3-distance(p) :
q 3-distance(q) :
Function k-distance(p): distance from p to the its k-nearest
neighbor
k-distance plot: k-distances of all objects, sorted in decreasing
order
DBSCAN: Heuristics for determining
EPS and MinPts
Idea is that for points in a cluster, their kth nearest
neighbors are at roughly the same distance
Noise points have the kth nearest neighbor at farther
distance
So, plot sorted distance of every point to its kth nearest
neighbor (e.g., k=4)
Thus, eps=10
26
Determining the Parameters
and MinPts
Example k-distance plot
3-distance
first „valley“
Objects
„border object“
Heuristic method:
Fix a value for MinPts (default: 2 d –1)
User selects “border object” o from the MinPts-distance plot;
is set to MinPts-distance(o)
Determining the Parameters
and MinPts
Problematic example
A C
F A, B, C
E
B, D, E
3-Distance
G
B‘, D‘, F, G
G1
G3 D1, D2,
D G2 G1, G2, G3
B D’
B’ D1
D2 Objects
Density Based Clustering:
Discussion
Advantages
Clusters can have arbitrary shape and size
Number of clusters is determined automatically
Can separate clusters from surrounding noise
Can be supported by spatial index structures
Disadvantages
Input parameters may be difficult to determine
In some situations very sensitive to input parameter
setting
OPTICS: Ordering Points To Identify
the Clustering Structure
DBSCAN
Input parameter – hard to determine.
Algorithm very sensitive to input parameters.
OPTICS – Ankerst, Breunig, Kriegel, and
Sander (SIGMOD’99)
Based on DBSCAN.
Does not produce clusters explicitly.
Rather generate an ordering of data objects
representing density-based clustering structure.
OPTICS con’t
Produces a special order of the database wrt its
density-based clustering structure
This cluster-ordering contains info equiv to the
density-based clusterings corresponding to a broad
range of parameter settings
Good for both automatic and interactive cluster
analysis, including finding intrinsic clustering
structure
Can be represented graphically or using
visualization techniques
Density-Based Hierarchical
Clustering
Observation: Dense clusters are completely contained
by less dense clusters
C D
C1 C2
Idea: Process objects in the “right” order and keep track of point
density in their neighborhood
C MinPts = 3
C1 C2
2 1
Core- and Reachability Distance
Parameters: “generating” distance fixed value
MinPts
core-distance,MinPts(o)
“smallest distance such that o is a core object”
(if that distance is “?”otherwise)
MinPts = 5
reachability-distance,MinPts(p, o)
“smallest distance such that p is p
directly density-reachable from o” q o
(if that distance is “?”otherwise)
core-distance(o)
reachability-distance(p,o)
reachability-distance(q,o)
OPTICS: Extension of
DBSCAN
Order points by shortest reachability distance to
guarantee that clusters w.r.t. higher density are finished
first. (for a constant MinPts, higher
density requires lower ε)
The Algorithm OPTICS
Basic data structure: controlList
Memorize shortest reachability distances seen so far
(“distance of a jump to that point”)
Visit each point
Make always a shortest jump
Output:
order of points
core-distance of points
reachability-distance of points
The Algorithm OPTICS
ControlList ordered by reachability-distance.
ControlList cluster-ordered
foreach o Database file
// initially, o.processed = false for all objects o
if o.processed = false;
insert (o, “?”) into ControlList;
while ControlList is not empty
database
select first element (o, r-dist) from ControlList;
retrieve N(o) and determine c_dist= core-distance(o);
set o.processed = true;
write (o, r_dist, c_dist) to file;
if o is a core object at any distance
foreach p N(o) not yet processed;
determine r_distp = reachability-distance(p, o);
if (p, _) ControlList
insert (p, r_distp) in ControlList;
else if (p, old_r_dist) ControlList and r_distp old_r_dist
update (p, r_distp) in ControlList;
OPTICS: Properties
“Flat” density-based clusters wrt. * andMinPts afterwards:
Starts with an object o where c-dist(o) * and r-dist(o) > *
Continues while r-dist *
1
2 17
3 16 18 34
1
4 2 16 17
18
Core-distance
Performance: approx. Reachability-distance
runtime( DBSCAN( , MinPts) )
O( n * runtime(-neighborhood-query) )
without spatial index support (worst case): O( n2 )
e.g. tree-based spatial index support: O( n log(n) )
OPTICS: The Reachability Plot
represents the density-based clustering
structure
easy to analyze
independent of the dimension of the data
reachability distance
reachability distance
cluster ordering cluster ordering
OPTICS: Parameter Sensitivity
Relatively insensitive to parameter settings
Good result if parameters are just 1 3
“large enough”
2
MinPts = 10, = 10 MinPts = 10, = 5 MinPts = 2, = 10
1 2 3
1 2 3
1 2 3
An Example of OPTICS
neighboring objects stay close to each other in a linear
sequence.
Reachability-
distance
undefined
‘
Cluster-order of the objects
DBSCAN VS OPTICS
DBSCAN OPTICS
Density Boolean value Numerical value
(high/low) (core distance)
Density- Boolean value Numerical value
connected (yes/no) (reachability distance)
Searching random greedy
strategy
When OPTICS Works Well
Cluster-order of the objects
When OPTICS Does NOT
Work Well
Cluster-order of the objects
DENCLUE: using density
functions
DENsity-based CLUstEring by Hinneburg & Keim
(KDD’98)
Major features
Solid mathematical foundation
Good for data sets with large amounts of noise
Allows a compact mathematical description of arbitrarily
shaped clusters in high-dimensional data sets
Significantly faster than existing algorithm (faster than
DBSCAN by a factor of up to 45)
But needs a large number of parameters
Denclue: Technical Essence
Model density by the notion of influence
Each data object exert influence on its neighborhood.
The influence decreases with distance
Example:
Consider each object is a radio, the closer you are to the
object, the louder the noise
Key: Influence is represented by mathematical
function
Denclue: Technical Essence
Influence functions: (influence of y on x, is a user given constant)
Square : f ysquare(x) = 0, if dist(x,y) > ,
1, otherwise
Guassian:
d ( x, y )2
y 2 2
f Gaussian ( x) e
Density Function
Density Definition is defined as the sum of the
influence functions of all data points.
d ( x , xi ) 2
( x ) i 1 e
D N
2 2
f Gaussian
Gradient: The steepness of a
slope
Example
d ( x , y )2
f Gaussian ( x , y ) e 2 2
d ( x , xi ) 2
( x ) i 1 e
D N
2 2
f Gaussian
d ( x , xi ) 2
( x, xi ) i 1 ( xi x) e
D N
2 2
f Gaussian
Denclue: Technical Essence
Clusters can be determined mathematically
by identifying density attractors.
Density attractors are local maximum of the
overall density function.
Density Attractor
Cluster Definition
Center-defined cluster
A subset of objects attracted by an attractor x
density(x) ≥
Arbitrary-shape cluster
A group of center-defined clusters which are
connected by a path P
For each object x on P, density(x) ≥ .
Center-Defined and Arbitrary
DENCLUE: How to find the
clusters
Divide the space into grids, with size 2
Consider only grids that are highly
populated
For each object, calculate its density
attractor using hill climbing technique
Tricks can be applied to avoid calculating
density attractor of all points
Density attractors form basis of all clusters
Features of DENCLUE
Major features
Solid mathematical foundation
Compact definition for density and cluster
Flexible for both center-defined clusters and arbitrary-
shape clusters
But needs a large number of parameters
: parameter to calculate density
: density threshold
: parameter to calculate attractor
