Skip to contents

Project new data onto an existing SOM + cluster solution

Source: R/Project_SOMClust.R
Project_SOMClust.Rd

Given a fitted Pipeline_SOMClust object, project a new data frame onto:

  • The same Z-score scaling (via SciDataReportR::Project_ZScore()) when the training object used computed or projected Z-scores.

  • The same SOM (kohonen), using kohonen::map().

  • The same node-level latent profiles and posterior probabilities.

If the training object used ZScoreType = "PreZScored", this function expects the new data already contains the same Z-score columns (by name) stored in object$ZScoreVars. No re-zscoring is performed.

For each projected case, the function:

  • Assigns a SOM node and distance to BMU.

  • Maps node-level cluster/posterior probabilities to the individual.

  • Computes an overall SOM-distance z-score and flags:

    • Flag_SOMDist_overallHigh – distance > overall training p95.

    • Flag_SOMDist_clusterHigh – distance > cluster-specific training p95 for that cluster.

Stable row id:

  • .scidr_rowid is added to new_df (if not already present) and carried into df_with_clusters and ProbFit$individual.

  • ProbFit$individual$RowID is set equal to .scidr_rowid.

Missing data:

  • Only rows with complete Z-scores are mapped to the SOM.

  • Rows with missing Z-scores receive NA for SOM_Node, SOM_Distance, and the cluster label.

Usage

Project_SOMClust(object, new_df, ClusterName = NULL)

Arguments

object

A Pipeline_SOMClust object from a previous run.

new_df

Data frame of new cases to project.

ClusterName

Optional name for the cluster column; defaults to object$ClusterName. If that column already exists in new_df it is overwritten (with a message).

Value

A list of class "Project_SOMClust" with components:

  • vars_used, ClusterName, complete_rows

  • df_with_clusters: new_df with .scidr_rowid and only the cluster column appended.

  • SOMProj: list with training and projected distance summaries, cluster-level flag summaries, comparison, and plots.

  • ProbFit: list with node (training node-level posterior info), individual (projection-level info including distance flags and z-scores), and probability plots.

  • ModelInfo_SOM, ModelInfo_MClust: references to the original model objects for convenience.