rstreet85/JKmeans

A Java program for clustering data with the k-means algorithm.

⚠️ This project is outdated, I just keep it here as a record of my grad work ⚠️

JKmeans

Java implementation of the k-means clustering algorithm. K-means accepts N d-dimensional data points and groups them into k clusters, choosing the nearest cluster based on the distance between the data point and the prototype (centroid) of the cluster. Simply, it is a way of grouping a large number of disparate points into a small number of similar groups, then finding the "ideal" member of each group. After grouping, the centroid of each cluster can be updated, and the clustering repeated over i iterations. After each iteration, the sum of each cluster's sum of squared errors (SSE) will be lower than the previous iteration. K-means is a simple minimum-seeking algorithm, and will stop at a local minimum, even if a better one exists - therefore the issue of how initial centroids are seeded becomes the most critical.

⚠️ K-means is not suited to every data clustering project. Membership is mutually exclusive - a point cannot be in more than one cluster at once. If any of the characteristics on which the points are being clustered are not useful in discriminating classes, the final clustering can be affected, so too much data can be an issue. ⚠️

Usage

Command-line: -fileName -numberClusters -numberIterations numberDataPoints numberDataDimensions

Data Format

Program assumes a valid CSV format, no header, no label data file from the UC-Irvine Machine Learning Repository.

To-Do

• Add a 'KmeansFactory' as wrapper for complex Kmeans initialization.
• Additional 'smart' seeding methods (k-means++).
• Additional distance measurements (Manhattan, Mahalanobis, etc.).
• Rand index to measure correlation.
• Add meaningful labels to output.
• Print output to file.
• Ability to cluster text data points (Ex via Hamming).
• Ability to cluster data points that have categorical (text/character/uid) attributes.