1. 2015/6/201 Minimum Spanning Tree Partitioning Algorithm for Microaggregation 報告者：林惠珍

2. 2015/6/202 Outline The Microaggregation problem MST partitioning algorithm Experimental results Conclusions and future work

3. 2015/6/203 The Microaggregation problem Microdata set( n records, p numerical attributes) To partition the n points into groups so as to minimize the objective function ： SSE subject to It is equivalent to minimizing a standardized information loss measure L= SSE / SST.

4. 2015/6/204 Microaggregation vs. clustering problem Constraint Problem The number of clusters cluster size MicroaggregationNoYes ClusteringYesNo So, modify the strategy for selecting edges for deletion from the MST for microaggregation problem. In decreasing length order Each tree has at least k nodes

5. 2015/6/205 MST Partitioning Algorithm MST construction( Prim’s algorithm) Edge cutting Cluster formation

6. 2015/6/206 k=5

7. 2015/6/207

8. 8 Experimental methods D: Diameter-based fixed size method C: Centroid-based fixed size method M: MST-partitioning algorithm M-d: MST-partitioning followed by clusters of size >= 2k partitioned by D M-c: MST-partitioning followed by clusters of size >= 2k partitioned by C

9. 2015/6/209

10. 10 Experimental results

11. 2015/6/2011 Conclusions The more pronounced the inherent clustering effects in the data, the greater is the advantage of using the our methods. MST partitioning-based method should be considered as a potential candidate for any practical application.

12. 2015/6/2012 Future work To adapt some of the ideas used to solve other clustering problems to this constrained version. To explore methods where minimum group size is treated as a soft constraint associated with a preference level.