../E16808-07.epub /> ../E16808-07.mobi />
See Also:Chapter 7, "Clustering"
Campos, M.M., Milenova, B.L., "Clustering Large Databases with Numeric and Nominal Values Using Orthogonal Projections", Oracle Data Mining Technologies, Oracle Corporation.
This chapter contains the following topics:
O-Cluster is a fast, scalable grid-based clustering algorithm well-suited for mining large, high-dimensional data sets. The algorithm can produce high quality clusters without relying on user-defined parameters.
The objective of O-Cluster is to identify areas of high density in the data and separate the dense areas into clusters. It uses axis-parallel uni-dimensional (orthogonal) data projections to identify the areas of density. The algorithm looks for splitting points that result in distinct clusters that do not overlap and are balanced in size.
O-Cluster operates recursively by creating a binary tree hierarchy. The number of leaf clusters is determined automatically. The algorithm can be configured to limit the maximum number of clusters.
Partitioning strategy refers to the process of discovering areas of density in the attribute histograms. The process differs for numerical and categorical data. When both are present in the data, the algorithm performs the searches separately and then compares the results.
In choosing a partition, the algorithm balances two objectives: finding well separated clusters, and creating clusters that are balanced in size. The following paragraphs detail how partitions for numerical and categorical attributes are identified.
To find the best valid cutting plane, O-Cluster searches the attribute histograms for bins of low density (valleys) between bins of high density (peaks). O-Cluster attempts to find a pair of peaks with a valley between them where the difference between the peak and valley histogram counts is statistically significant.
A sensitivity level parameter specifies the lowest density that may be considered a peak. Sensitivity is an optional parameter for numeric data. It may be used to filter the splitting point candidates.
Categorical values do not have an intrinsic order associated with them. Therefore it is impossible to apply the notion of histogram peaks and valleys that is used to partition numerical values.
Instead the counts of individual values form a histogram. Bins with large counts are interpreted as regions with high density. The clustering objective is to separate these high-density areas and effectively decrease the entropy (randomness) of the data.
O-Cluster identifies the histogram with highest entropy along the individual projections. Entropy is measured as the number of bins above sensitivity level. O-Cluster places the two largest bins into separate partitions, thereby creating a splitting predicate. The remainder of the bins are assigned randomly to the two resulting partitions.
After processing an initial random sample, O-Cluster identifies cases that are of no further interest. Such cases belong to frozen partitions where further splitting is highly unlikely. These cases are replaced with examples from ambiguous regions where further information (additional cases) is needed to find good splitting planes and continue partitioning. A partition is considered ambiguous if a valid split can only be found at a lower confidence level.
Cases associated with frozen partitions are marked for deletion from the buffer. They are replaced with cases belonging to ambiguous partitions. The histograms of the ambiguous partitions are updated and splitting points are reevaluated.
The O-Cluster algorithm evaluates possible splitting points for all projections in a partition, selects the best one, and splits the data into two new partitions. The algorithm proceeds by searching for good cutting planes inside the newly created partitions. Thus O-Cluster creates a binary tree structure that divides the input space into rectangular regions with no overlaps or gaps.
The main processing stages are:
Load the buffer. Assign all cases from the initial buffer to a single active root partition.
Compute histograms along the orthogonal uni-dimensional projections for each active partition.
Find the best splitting points for active partitions.
Flag ambiguous and frozen partitions.
When a valid separator exists, split the active partition into two new active partitions and start over at step 2.
Reload the buffer after all recursive partitioning on the current buffer is completed. Continue loading the buffer until either the buffer is filled again, or the end of the data set is reached, or until the number of cases is equal to the data buffer size.
Note:O-Cluster requires at most one pass through the data
The clusters discovered by O-Cluster are used to generate a Bayesian probability model that can be used to score new data. The generated probability model is a mixture model where the mixture components are represented by a product of independent normal distributions for numerical attributes and multinomial distributions for categorical attributes.
The O-Cluster algorithm supports two build-time settings. Both settings have default values. There is no reason to override the defaults unless you want to influence the behavior of the algorithm in some specific way.
You can configure O-Cluster by specifying any of the following:
Buffer size — Size of the processing buffer. (See Active Sampling.)
Sensitivity factor — A fraction that specifies the peak density required for separating a new cluster. (See Partitioning Strategy.)
See Also:Oracle Database PL/SQL Packages and Types Reference for details about the build settings for O-Cluster
Automatic Data Preparation bins numerical attributes for O-Cluster. It uses a specialized form of equi-width binning that computes the number of bins per attribute automatically. Numerical columns with all nulls or a single value are removed.
O-Cluster handles missing values naturally as missing at random. The algorithm does not support nested columns and thus does not support sparse data.
Oracle Data Mining Application Developer's Guide for information about nested columns and missing data
Keep the following in mind if you choose to prepare the data for O-Cluster:
O-Cluster does not necessarily use all the input data when it builds a model. It reads the data in batches (the default batch size is 50000). It will only read another batch if it believes, based on statistical tests, that there may still exist clusters that it has not yet uncovered.
Binary attributes should be declared as categorical.
Automatic equi-width binning is highly recommended. The bin identifiers are expected to be positive consecutive integers starting at 1. See Oracle Database PL/SQL Packages and Types Reference for an example.
The presence of outliers can significantly impact clustering algorithms. Use a clipping transformation before binning or normalizing. Outliers with equi-width binning can prevent O-Cluster from detecting clusters. As a result, the whole population appears to fall within a single cluster.