ELKI MiniGUI中K-Means聚类如何指定CSV特征列？

Got it, let's tackle your question about ELKI and K-Means feature selection.

Short answer: No, the ELKI MiniGUI doesn't support selecting specific feature columns directly. By default, it uses all numeric columns in your CSV for clustering, and ignores non-numeric columns (so if your label column is a string, it'll skip it automatically). But if you want to pick a subset of your 15 feature columns, the MiniGUI doesn't have a built-in option for that—you'll need to either preprocess your data (create a CSV with only the columns you want) or use the Java/command-line version of ELKI.

The easiest approach doesn't require modifying ELKI's source code at all—you can use ELKI's built-in SelectedFeatureFilter to pick exactly the columns you want. Here are two methods:

Option 1: Use the ELKI Command Line (No Coding Needed) #

This is the quickest way if you don't want to write Java code. Just run ELKI with the following command (adjust paths and parameters to your dataset):

java -jar elki.jar KDDCLIApplication \
  -dbc.in /path/to/your/dataset.csv \
  -dbc.parser NumberVectorLabelParser \
  -dbc.parser.labelcolumn 15 \  # Index of your label column (0-based)
  -filter SelectedFeatureFilter \
  -filter.selectedfeatures 0,2,5-7 \  # Your desired feature column indices (comma-separated, ranges allowed)
  -algorithm clustering.kmeans.KMeansLloyd \
  -kmeans.k 5 \  # Number of clusters you want
  -distancefunction SquaredEuclideanDistanceFunction

Breakdown:

• -dbc.parser.labelcolumn 15 tells ELKI to treat column 15 (0-based) as the label (so it won't use it for clustering)
• -filter.selectedfeatures lets you specify exactly which feature columns to include—use commas for individual indices, or hyphens for ranges (e.g., 5-7 means columns 5,6,7)

Option 2: Write a Minimal Java Program #

If you prefer a programmatic approach (e.g., for automating multiple runs with different column combinations), here's a stripped-down example:

import de.lmu.ifi.dbs.elki.data.DoubleVector;
import de.lmu.ifi.dbs.elki.database.Database;
import de.lmu.ifi.dbs.elki.database.connection.FileBasedDatabaseConnection;
import de.lmu.ifi.dbs.elki.data.filter.selection.SelectedFeatureFilter;
import de.lmu.ifi.dbs.elki.datasource.parser.CSVReaderFormat;
import de.lmu.ifi.dbs.elki.datasource.parser.NumberVectorLabelParser;
import de.lmu.ifi.dbs.elki.algorithm.clustering.kmeans.KMeansLloyd;
import de.lmu.ifi.dbs.elki.distance.distancefunction.minkowski.SquaredEuclideanDistanceFunction;
import de.lmu.ifi.dbs.elki.algorithm.clustering.kmeans.initialization.RandomlyGeneratedInitialMeans;

public class KMeansFeatureSelection {
    public static void main(String[] args) {
        // Configure CSV parsing: label column is index 15 (0-based)
        CSVReaderFormat csvFormat = new CSVReaderFormat()
                .withParser(new NumberVectorLabelParser(DoubleVector.FACTORY, 15));

        // Set up database with feature selection filter
        FileBasedDatabaseConnection dbConnection = new FileBasedDatabaseConnection(
                "/path/to/your/dataset.csv",
                csvFormat,
                // Select specific feature columns: e.g., 0,1,3,4
                new SelectedFeatureFilter(new int[]{0,1,3,4})
        );

        // Initialize database
        Database db = dbConnection.initialize();

        // Configure K-Means
        KMeansLloyd<DoubleVector> kmeans = new KMeansLloyd<>(
                SquaredEuclideanDistanceFunction.STATIC,
                5,          // Number of clusters
                100,        // Max iterations
                new RandomlyGeneratedInitialMeans()
        );

        // Run and print results
        kmeans.run(db).output();
    }
}

You can easily loop over different column combinations in this code to automate multiple runs—just change the array in SelectedFeatureFilter each time.

• ELKI uses 0-based column indices, so make sure you count correctly (your 15 feature columns are probably indices 0-14, with label at 15)
• If you need to normalize/scale features, you can add additional filters (like NormalizeFilter) before the SelectedFeatureFilter
• No need to modify ELKI's source code—all the tools you need are already built into the library

内容的提问来源于stack exchange，提问作者gregory