On Using Class-Labels in Evaluation of Clusterings

Transcription

1 On Using Class-Labels in Evaluation of Clusterings Ines Färber Stephan Günnemann Hans-Peter Kriegel Peer Kröger Emmanuel Müller Erich Schubert Thomas Seidl Arthur Zimek RWTH Aachen University, Germany LMU Munich University, Germany MultiClust at KDD 2010 July 25, 2010

2 The Dilemma of Evaluation What would be the optimal clustering solution? View 1 View 2 On Using Class-Labels in Evaluation of Clusterings 1 / 1

3 Introduction evaluation of clustering solutions: evaluation based on internal measures + no additional information needed; data independent - approaches optimizing the evaluation criteria will always be preferred evaluation based on an experts opinion + may reveal new insight into the data - very expensive; results are not comparable evaluation based on external measures + objective evaluation - needs a valid ground truth On Using Class-Labels in Evaluation of Clusterings 2 / 1

4 History of Cluster Evaluation clustering broke off from classification assumption: classes stand out by inherent similarity traditional clustering mainly follows the partitioning approach external evaluation of traditional clustering: the original assumption motivated the comparison against class labels UCI - iris dataset class structure does not necessarily correspond to a clustering structure classes may split up into several subgroups there might be smooth transitions between two classes On Using Class-Labels in Evaluation of Clusterings 3 / 1

5 Multi-View Context assumption: data groups differently when seen from different perspectives each object might be grouped in multiple clusters with each perspective a set of attributes can be associated View 1 View 2 clustering goes beyond the structure of class labels data items potentially belong to many clusters in differing views class labels do not meet the assumptions of this scenario On Using Class-Labels in Evaluation of Clusterings 4 / 1

6 Classes vs. Clusters commonly observed differences between clusterings and class labelings: splitting of classes into multiple clusters merging of classes into a single cluster missing class outliers multiple (overlapping) hidden structures given class label C = shape alternative labels H = color Y W X Z On Using Class-Labels in Evaluation of Clusterings 5 / 1

7 Case Study Pendigits Dataset differnet ways of digit notation different types of digits 9 and 3 almost 30 different groups of digits in contrast to 10 given classes On Using Class-Labels in Evaluation of Clusterings 6 / 1

8 Case Study ALOI Dataset object groups that stand out due to similarity based on: color shape rotation object types feature space influences the clustering result On Using Class-Labels in Evaluation of Clusterings 7/1

9 Can Anything Be Learned? Findin QUESTIONABLE EDSC still widely used! On Using Class-Labels in Evaluation of Clusterings 8 / 1

10 Challenges 1 ground truth should provide multiple labellings 2 measures should be able to deal with multiple labels e.g. label layers challenges: clustering covers only part of the layer (incompleteness?) clusters in one layer vs. multiple layers (purity vs. variety) the clustering intersects layers the clustering contains newly detected clusters e.g. label hierarchies e.g. label ontologies On Using Class-Labels in Evaluation of Clusterings 9 / 1

11 Challenges 1 ground truth should provide multiple labellings 2 measures should be able to deal with multiple labels e.g. label layers e.g. label hierarchies challenges: might be hard to derive clustering covers one branch (redundancy?) clustering covers one layer (impurity?) clustering covers nodes only partially (incompleteness?) union of nodes newly detected clusters e.g. label ontologies On Using Class-Labels in Evaluation of Clusterings 10 / 1

12 Conclusion classification data class label per object database C H 1 hidden clusters per object H 2 H 3 H 4... clustering evaluation enhanced evaluation result proceed in the development of new clustering algorithms ensure objective clustering evaluation labeling of data measures for multiple labels On Using Class-Labels in Evaluation of Clusterings 11 / 1