Relevant Subspace Clustering:
Mining the Most Interesting Non-Redundant Concepts in High Dimensional Data

In. Proc. IEEE International Conference on Data Mining (ICDM 2009), Miami, USA (2009)

Emmanuel Müller, Ira Assent, Stephan Günnemann, Ralph Krieger and Thomas Seidl

Supplementary material concerning repeatability

On this website we provide supplementary matirial enabling repeatability of our experiments. Please refer to our paper for further details about our novel clustering model and the underlying algorithm. We focus on this website only on the used parameter settings and data sets for our evaluation which is based on the implementation in the OpenSubspace framework.

Using the WEKA framework for repeatability  

For an easy repeatability we integrated all algorithms for clustering in subspace projections of high dimensional data into the popular WEKA framework. We therefore extended the framework to subspace clustering. A short description of this extension and how to use it can be found on our OpenSubspace website. It includes a video tutorial giving a short introduction. Using our framework one can perform all evaluation measurements presented in our paper. Furthermore, one can interactively explore the clustering results.

Repeatability results

We expect repeatability results to be reasonably similar. There will be made two observations, one by us as the authors of this paper and one by other researchers repeating our experiments. These two observations should be similar; however they also might slightly varying from each other. For instance, there is no way to ensure that our hardware used in the experiments is available to other researchers repeating the evaluation. Therefore, we do not expect measured execution times to match those reported in the paper, but curve tendencies should be the same. When measuring other things than running time, such as e.g. result sizes in experiments with no randomized component, we do expect to obtain the results presented in the paper. Some of the approaches, however, include randomized components and thus will only show in the average case similar results, single runs might vary. Randomized components are e.g. random initialization of cluster centers in the PROCLUS algorithm.

Parameter settings

For repeatability, names of parameters are as in the original publications. Please refer to the original publications for a more detailed description. Some parameters have not been described or named in the publications; therefore we tried to give them as meaningful names as possible. Most of these not further described parameters are included as we use the original implementations provided by the authors of INSCY and FIRES. For example, FIRES has its main parameters (K, MINCLU and MU) while pre- and post-processing parameters are only of minor interest. For our relevant subspace clustering approach, we have given additionally to the main model parameters DELTA and BETA further parameters due to the cluster instantiation for density-based clustering and algorithmic parameters for grid size, merge operations on grids and pruning in our ranking. We have optimized all parameters for each algorithm on each data set and listed them for repeatability on this website.

Citation Information

If you publish results based on our material, then please include a reference to our paper. This will help others to obtain the same data sets, algorithms, parameter settings and evaluation measures and replicate your experiments. We suggest the following reference format for referring to this paper:

Müller E., Assent I., Günnemann S., Krieger R., Seidl T.:
Relevant Subspace Clustering: Mining the Most Interesting Non-Redundant Concepts in High Dimensional Data
In Proc. IEEE International Conference on Data Mining (ICDM 2009), Miami, USA. (2009) 

Parameter Settings for repeatability of our experiments:

Figure 5 (a): Result size w.r.t. dimensionality of the database




Figure 5 (b): Runtime w.r.t. dimensionality of the database

Figure 5 (c): Runtime w.r.t. size of the database

Figure 6. Quality on pendigits data (7494 objects; 16-48 dimensions)

Figure 7: F1 and accuracy on real data
[Captions: data set (size; dimensionality)]

Figure 8: Approximation quality (delta variation)