An Empirical Comparison of Stream Clustering Algorithms

Transcription

1 MÜNSTER An Empirical Comparison of Stream Clustering Algorithms Matthias Carnein Dennis Assenmacher Heike Trautmann CF 17 BigDAW Workshop Siena Italy May

2 Clustering MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 2 / Clustering aims to discover groups of similar objects Traditional algorithms assume a fixed set of data If new data arrives: Repeat the entire procedure

3 Stream Clustering MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 3 /29 6 data stream Stream Model 1. Data points arrive as a continuous stream 2. The size of the stream is large and potentially unbounded 3. Data points can only be evaluated once 4. The order of data points cannot be influenced 5. The distribution of the data can change over time

4 Outline MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 4 /29 1. Stream Clustering 2. Setup 3. Results 4. Conclusion

5 Outline MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 4 /29 1. Stream Clustering 2. Setup 3. Results 4. Conclusion

6 MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 5 /29 Two-Stage Clustering data stream online micro-clusters offline Reclustering macro-clusters [Aggarwal et al. 23]

7 Distance-Based Grid-Based BIRCH [Zhang et al. 1996] STREAM [O Callaghan et al. 22] CluStream [Aggarwal et al. 23] DenStream [Cao et al. 26] ClusTree [Kranen et al. 29] streamkm++ [Ackermann et al. 212] BICO [Fichtenberger et al. 213] DBSTREAM [Hahsler et al. 216] D-Stream [Chen et al. 27] D-Stream + attraction [Tu et al. 29] Challenges Many parameters No clear guidelines how to set them

8 MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 7 /29 DenStream #points linear sum per dimension squared sum per dimension Clustering Feature CF = ( n LS SS ) describes micro-cluster Decay the weight of clusters over time: ω(t) = 2 λt Core clusters can decay to outlier clusters attempt 3 attempt 1 attempt 2 outlier micro-cluster core micro-cluster

9 MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 8 /29 D-Stream Maintains density of grid cells Reclustering: combine neighbouring dense cells Extension using attraction: only combine cells that share points at the border

10 MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 9 /29 DBSTREAM Assign points based on distance threshold Move clusters towards new points (competitive learning) Maintains shared density between micro-clusters Reclustering: Merge clusters with high shared density

11 MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 1 /29 Remaining Algorithms BIRCH: maintain Clustering Features in a tree structure STREAM: repeatedly apply k-median approximation on chunks of data CluStream: store latest snapshots for various time intervals to extract portion of entire stream streamkm++: divisive clustering approach using medoids BICO: similar tree structure as BIRCH but with medoids similar to streamkm++ Research Goal Fair evaluation and comparison of ten popular stream clustering algorithms

12 Outline MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 11 /29 1. Stream Clustering 2. Setup 3. Results 4. Conclusion

13 W ESTFÄLISCHE W ILHELMS -U NIVERSITÄT M ÜNSTER An Empirical Comparison of Stream Clustering Algorithms 12 /29 Data sets Description square1 spiral complex9 t8.8k KDD Cup'99 Sensor Covertype Spherical Connected Various shapes Shapes + Noise Network Intrusion Environment sensors Forest type Stream Clustering 5 Data set Matthias Carnein Observations Setup 24 6 Dimension Cluster Noise % % % 4% Results 5 1 Conclusion

14 Evaluation MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 13 /29 Static Data: Evaluate clusters at the end of the stream Real world Data: Prequential evaluation [Cao et al. 26] 1. Evaluate 2. Update 3. Evaluate 4. Update 5. Evaluate 6. Update External Evaluation Micro-clusters: Purity Proportion of points that belong to the majority class in a cluster Macro-clusters: Adjusted Rand Index (ARI) Similarity between true groups and generated clusters

15 Parameter-Tuning MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 14 /29 Choice of parameters is important for a fair comparison Automatic parameter-configuration using Iterated Racing (irace) [López-Ibáñez et al. 216] Search for configurations that optimize the ARI of the macro-clustering Upper limit on the number of micro-clusters Static data: 5% of the number of observations in the stream Real world data: fixed upper limit of 5

16 Outline MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 15 /29 1. Stream Clustering 2. Setup 3. Results 4. Conclusion

17 Quality square1 BIRCH STREAM CluStream DenStream D-Stream D-Stream + attr. ClusTree StreamKM++ BICO DBSTREAM Purity 1.5 ARI 1.5 Micro-clusters: generally high purity inconsistency for DenStream Macro-clusters: good results but lowest ARI for DBSTREAM Example: Clustering result of DBSTREAM

18 Quality spiral BIRCH STREAM CluStream DenStream D-Stream D-Stream + attr. ClusTree StreamKM++ BICO DBSTREAM 1 1 Purity.5 ARI.5 Micro-clusters: high purity for CluStream DBSTREAM BICO streamkm++ Macro-clusters: only decent results for DBSTREAM Example: clustering result of STREAM

19 Quality Sensor BIRCH STREAM CluStream DenStream D-Stream D-Stream + attr. ClusTree StreamKM++ BICO DBSTREAM 1 Purity ARI Time (in 1) Time (in 1)

20 Runtime Various programming languages (R C C++ Java) Different interfaces different authors BIRCH STREAM CluStream DenStream D-Stream D-Stream + attr. ClusTree StreamKM++ BICO DBSTREAM Runtime (in s) square1 spiral complex9 t8.8k KDD Cup'99 Sensor Covertype

21 Outline MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 2 /29 1. Stream Clustering 2. Setup 3. Results 4. Conclusion

22 Conclusion MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 21 /29 Empirical comparison of 1 stream clustering algorithms DBSTREAM performed best but dependent on insertion order D-Stream good results but requires more micro-cluster DenStream poor results and very inconsistent Future Work More algorithms and data sets and runtime experiments Alternative configuration methods Use results to design an improved algorithm Application to Multi-Channel Customer Relationship Management

23 MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 22 /29 Questions??????

24 References I MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 23 /29 Ackermann Marcel R. Marcus Märtens Christoph Raupach Kamil Swierkot Christiane Lammersen and Christian Sohler (212). StreamKM++: A Clustering Algorithm for Data Streams. In: J. Exp. Algorithmics : :2.3. (Visited on 16/2/217). Aggarwal Charu C. Jiawei Han Jianyong Wang and Philip S. Yu (23). A Framework for Clustering Evolving Data Streams. In: Proceedings of the 29th International Conference on Very Large Data Bases. Vol. 29. VLDB 3. Berlin Germany: VLDB Endowment pp Cao Feng Martin Ester Weining Qian and Aoying Zhou (26). Density-based clustering over an evolving data stream with noise. In: Conference on Data Mining (SIAM 6) pp DOI: 1.119/CCCM Chen Yixin and Li Tu (27). Density-based Clustering for Real-time Stream Data. In: Proceedings of the 13th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. KDD 7. San Jose California USA: ACM pp DOI: /

25 References II MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 24 /29 Fichtenberger Hendrik Marc Gillé Melanie Schmidt Chris Schwiegelshohn and Christian Sohler (213). BICO: BIRCH Meets Coresets for k-means Clustering. In: Algorithms - ESA st Annual European Symposium Sophia Antipolis France September Proceedings. URL: (visited on 2/3/217) pp DOI: 1.17/ _41. Guha S. A. Meyerson N. Mishra R. Motwani and L. O Callaghan (23). Clustering data streams: Theory and practice. In: IEEE Transactions on Knowledge and Data Engineering 15.3 pp DOI: 1.119/TKDE Hahsler Michael and Matthew Bolaños (216). Clustering Data Streams Based on Shared Density between Micro-Clusters. In: IEEE Transactions on Knowledge and Data Engineering 28.6 pp DOI: 1.119/TKDE Kranen Philipp Ira Assent Corinna Baldauf and Thomas Seidl (29). Self-Adaptive Anytime Stream Clustering. In: Ninth IEEE International Conference on Data Mining (ICDM 9) pp DOI: 1.119/ICDM

26 References III MÜNSTER An Empirical Comparison of Stream Clustering Algorithms 25 /29 López-Ibáñez Manuel Jérémie Dubois-Lacoste Leslie Pérez Cáceres Thomas Stützle and Mauro Birattari (216). The irace package: Iterated Racing for Automatic Algorithm Configuration. In: Operations Research Perspectives. DOI: 1.116/j.orp O Callaghan L. N. Mishra A. Meyerson S. Guha and R. Motwani (22). Streaming-data algorithms for high-quality clustering. In: Proceedings of the 18th International Conference on Data Engineering (ICDE). URL: (visited on 2/3/217) pp DOI: 1.119/ICDE Tu Li and Yixin Chen (29). Stream Data Clustering Based on Grid Density and Attraction. In: ACM Trans. Knowl. Discov. Data :1 12:27. DOI: / Zhang Tian Raghu Ramakrishnan and Miron Livny (1996). BIRCH: An Efficient Data Clustering Method for Very Large Databases. In: Proceedings of the 1996 ACM SIGMOD International Conference on Management of Data. SIGMOD 96. New York NY USA: ACM pp DOI: /