Accepted paper at ICDM 2020 Workshop HDM'20
Daniyal Kazempour, Anna Beer, Peer Kröger, Thomas Seidl
The 8th ICDM Workshop on High Dimensional Data Mining (HDM'20)
in conjunction with the 20th IEEE International Conference on Data Mining (ICDM 2020),
17–20 November 2020, Virtual
In this work we propose SRE, the first internal evaluation measure for arbitrary oriented subspace clustering results. For this purpose we present a new perspective on the subspace clustering task: the goal we formalize is to compute a clustering which represents the original dataset by minimizing the reconstruction loss from the obtained subspaces, while at the same time minimizing the dimensionality as well as the number of clusters. A fundamental feature of our approach is that it is model-agnostic, i.e., it is independent of the characteristics of any specific subspace clustering method. It is scale invariant and mathematically founded. The experiments show that the SRE scoring better assesses the quality of an arbitrarily oriented subspace clustering compared to commonly used external evaluation measures.