ELKI 0.5.0 release notes
Primary release goals
- Cluster evaluation: metrics and circle-segment-visualization (ICDE 2012, see Examples/ClusterEvaluation)
- Outlier detection ensembles (SDM 2011, 2012, see Examples/GreedyEnsemble)
- Usability improvements, for example by adding an automatic evaluation helper
- Performance improvements by reducing boxing of primitive types (see DBID API for details)
- Parallel coordinates visualizations added for high-dimensional data
Additional improvements
Algorithms
Distance functions
Index layer
Evaluation
- Most clustering similarity measures (BCubed, Rand, ARI, Mutual Information, Entropy, Edit, …)
- More outlier evaluation measures (ROC, P/R curves, Average-Precision curves, SmROC curves)
- Automatic Evaluation, disable via
-evaluator NoAutomaticEvaluation
.
Visualizations
- Alpha shapes
- Voronoi cells for 2D K-means
- Cluster stars for mean-models (k-means, EM)
- Parallel coordinates
Applications
- Cluster similarity visualization, from ICDE 2012 (see Examples/ClusterEvaluation)
- Greedy Ensemble for Outlier Detection, from SDM 2012 (see Examples/GreedyEnsemble)
Other
- Improved support for sparse vectors. Note the need to apply the SparseVectorFieldFilter for many algorithms (that assume a fixed dimensionality vector field! For more details, see HowTo/SparseData
- Improved error logging (less exceptions lost, shorter stacktraces)
- Robustness improvements
- Use GNU Trove primitive collections for further speedups.
- DBID API changes, for performance improvements.
- No longer include PDF/PS/EPS support in
elki.jar
(add Apache FOP to your classpath to enable!)