clustering: DBSCAN Clustering on Each Study Wave

In unmnn/evoxploit: Augment Evolution Features to a Longitudinal Study Dataset

Description

Performs feature selection and DBSCAN clustering on each study wave independently.

Usage

clustering(df, label, minPts = NULL, eps = NULL, suffix = "_s", ...)

Arguments

df	The data frame. TODO: Add requirements, e.g. w.r.t. suffix
label	The class labels given as factor vector.
minPts	minPts. If not set, it defaults to log(n), where n is the number of objects, as suggested in \insertCiteKailing:RIS2003evoxploit.
eps	ε.

Details

The function first employs CFS \insertCiteHall:CFS2000evoxploit to obtain a small set of relevant, non-redundant features for DBSCAN clustering. If for a variable only one wave is selected (e.g. only som_bmi_s0 but not som_bmi_s1 and *._s2), it expands the feature space to include all possible realizations.

Then, for each wave, it applies DBSCAN. If neccessary, it uses a heuristic to find appropriate values for DBSCAN's parameters.

Distance calculation is based on heom. If no values for minPts and ε are provided, the parameter are estimated using a heuristic.

Value

A list containing the following elements for each wave:

• dist: The dist matrix used for clustering.

• subset_att_wave: A character vector of attribute names used for clustering.

• kdist: A list containing at least minPts and ε.

• clustering_result: The DBSCAN clustering output.

References

\insertAllCited

See Also

heom, best_att_subset_global, kdist_info

unmnn/evoxploit documentation built on Oct. 28, 2020, 12:24 p.m.