Dimension editor

• "Alphabetically" uses whatever order you've selected for the "Refinements sort order" setting on the main part of the Dimension editor.
• "Dynamically" orders the most popular refinement values according to their frequency of appearance within a data set. Dimension values that occur more frequently are returned before those that occur less frequently.

Specifies whether Cluster Discovery is enabled for this dimension. The checkbox also makes the following controls available.

For more information, see the Endeca Relationship Discovery Guide.

This parameter governs how many documents are sampled. Clustering processing time and memory consumption are both roughly linear with this number; thus, lowering the value results in smaller memory consumption and faster turnaround. However, statistical errors are likely to occur when the sample size is small. Setting this value higher will overcome statistical errors for data sets where fewer terms are tagged onto each document.

Range: Integer, 50-2000 (default: 500)

Recommended value: 500

Range: Integer, 2-10 (default: 10)

Recommended value: 6

This parameter governs the decision of whether a set of terms is coherent enough to form a cluster (that is, each cluster should have only closely related documents). Low values are permissive (i.e., not demanding much coherence) and will result in fewer larger clusters; high values are strict and will result in more smaller clusters. The average value is recommended.

Range: Integer, 2-10 (default: 10)

Recommended value: 6

Terms that are extracted from sampled documents are filtered by their precision p (where p = number of sampled documents that this term is tagged onto divided by the number of all sampled documents). Terms that have too high a value of p are likely to be the search term (or be synonymous with it) or be too general to make for a good clustering term. If you use the recommended tuning values of the term extractor, each term is tagged only roughly 1/3 of the documents, which means that the search term, if present, will have p of roughly 0.33 (more or less stringent tuning of the term extractor will change this value). There usually is a gap in the values of p between the search term and the more useful terms, which start at approximately p = 0.25 and less.

Range: Float, 0.0 - 1.0 (default: 1.0)

Recommended value: 0.25

Sets the maximum number of terms that can be in a cluster. Each cluster will have at least 2 terms. Because of the match-partial cluster selection mechanism, the more terms there are in the cluster, the (potentially) higher its coverage will be. On the other hand, the clusters that are too large take up too much space to display and take too long for users to read.

Range: Integer, 2 - 10 (default: 10)

Recommended value: 8

If two clusters overlap (that is, if the document sets that each cluster maps to overlap), then the smaller one (as measured by the estimated size of the document set it maps to) can be removed, depending on how big this overlap is. This parameter dictates the overlap above which the smaller cluster is removed.

Clusters which overlap by more than this value will be removed. Thus, the default setting of 10 (out of ten) means that clusters that overlap by more than 10 out of 10 documents will be removed. Since this is impossible, this means that setting of 10 will disable cluster overlap filtering, which is most extreme level of coarseness for this filter. Tuning this parameter down will make the cluster overlap more and more fine-grained. Thus, a value of 9 will remove only the clusters that greatly overlap; setting it to the recommended value of 5 will remove only clusters overlapping half-way or so (remember that the overlap is merely estimated). Setting this parameter to lower values (less than 5) will make overlap filtering quite sensitive and will remove clusters which overlap even by a small amount. Note that clusters that do not overlap at all will never be filtered.

Range: Integer, 0-10 (default: 10)

Recommended value: 5