1. So Hirai The University of Tokyo Currently NTT DATA Corp. Kenji Yamanishi The University of Tokyo WITMSE 2012, Amsterdam, Netherland Presented at KDD 2012 on Aug.13.

2. Contents Problem Setting Significance Proposed Algorithm ： Sequential Dynamic Model Selection with NML(normalized maximum likelihood) coding How to compute the NML coding for Gaussian mixtures Experimental Results Marketing Applications Conclusion 2

3. Problem Setting (1/2) 3 Time Change Clustering change detection ---Tracking changes of clustering structures in a sequential setting to detect novelty in data Ex. Market analysis The structure of customer groups changes over time Detect changes of the number of clusters as well as their assignment

4. Problem Setting (2/2) 4 F E DC B A F E D C B A F E D C B A F E D C B A α β Examples of clustering structure changes Existing customers change their patterns New customer s emerge to form a new group There exist various types of clustering structures

5. Related works Evolutionally clustering [Chakrabrti et. al., 2006] Hypothesis testing approach[Song and Wang, 2005] Kalman filter approach [Krempl et. al., 2011] Graph Scope [Sun et. al., 2007] Variational Bayes approach[Sato, 2001] 5 Clustering change detection issue

6. Significance A novel clustering change detection algorithm Key idea: ・ Sequential dynamic model selection (sequential DMS ） ・ NML(normalized maximum likelihood) code-length as criteria ……..First formulae for NML for Gaussian mixture models 6 Empirical demonstration of its superiority over existing methods Shown using artificial data sets Demonstration of its validity in market analysis Shown using real beer consumption data sets

7. Sequential Dynamic Model Selection Algorithm 7

8. Proposed Alg. – background of DMS – Batch DMS criterion : 8 Dynamic Model Selection ( DMS ) [Yamanishi and Maruyama, 2007] Total code-length Code-length of data seq. Code-length of model seq. Minimum w.r.t. ~Extension of MDL (Minimum Description Length) principle[Rissanen, 1978] into model “sequence” selection

9. Proposed Alg. – Sequential DMS – At each time t, given, sequentially select for clustering 9 Sequential dynamic model selection (SDMS) Alg. Code-length for data clustering ～ NML (normalized maximum likelihoood) coding Code-length for transition of clustering structure Minimum w.r.t. K t, Z t Sequential variant of DMS criterion [Yamanishi and Maruyama, 2007] s.t.

10. Proposed Alg. – model transition – Run EM alg. with initial values below: Case 1 # of clusters does not change Initial parameter values remain the same Case 2 # of clusters decreases (e.g., merging) Assign data in a certain cluster to other ones randomly Case 3 # of clusters increases (e.g., splitting) Set data to a new cluster randomly 10 Consider three patterns of clustering changes Case 2 Case 3

11. Proposed Alg. – code-length for transition – Model transition probability distribution Suppose K transits to neighbors only Employ Krichevsky-Trofimov (KT) estimate [Krichevsky and Trofimov, 1981] 11 Code-length of the model transition

12. How to compute NML code-length for Gaussian mixtures 12

13. Criteria – NML code-length – Model (Gaussian mixture model) : NML (normalized maximum likelihood) code-length : Shortest code-length in the sense of minimax criterion [Shatarkov 1987] 13 Normalization term

14. For Continuous Data Normalization term In case of, the data ranges over all domains Problem: NML for Gaussian distribution Normalization term diverges NML for mixture distribution Normalization term is computationally intractable This comes from combinational difficulties 14

15. For Continuous Data (Example) For the one-dimension Gaussian distribution (σ 2 is given) Normalization term 15

16. Approximate computation (1/2) 16 Use sufficient statistics g 1 : Gaussian distribution g 2 : Wishart distribution

17. Criteria – NML for GMM – Restrict the range of data so that the MLE lies in a bounded range specified by a parameter 17 Efficiently computing an approximate variant of the NML code-length for a GMM [Hirai and Yamanishi, 2011] The normalization term does not diverge But still highly depends on the parameters :

18. ＮＭＬ The normalization term is calculated as follows : 18 where, : number of data,: dim. of data

19. Criteria – RNML code-length – Re-normalize around the MLE of parameter by restricting the range of data 19 Modify NML to develop the re-normalized maximum likelihood coding (RNML) [Rissanen, Roos, Myllymaki 2010] [Hirai and Yamanishi, 2012] Less dependent on hyper-parameter

20. 20 Criteria – RNML code-length –

21. RNML code-length Theorem [Hirai and Yamanishi 2012] ＲＮＭＬ code-length for GMM is calculated as follows : 21 Definition Problem Computing, costs. 1

22. Criteria – efficient computing of RNML – Straightforward computation of RNML requires time ⇒ But we can compute it efficiently Theorem [Kontkanen and Myllymaki, 07] 22 １ )

23. Can compute the normalization term in for “mixture” models Criteria – efficient computing of RNML – Straightforward computation of RNML requires time ⇒ But we can compute it efficiently Theorem [Hirai and Yamanishi, 2012] The normalization term satisfies recurrsive formula 23 ２ ２ ２

24. Experimental Results – Artificial Data – – Market Analysis – 24

25. Experimental Results – data generation – Generate artificial data set according to GMM with 25

26. Experimental Results – comparison criteria – AR (accuracy rate) : Average rate of correctly estimating the true number of clusters over all time IR (identification rate) : Probability of correctly identifying change-points and change themselves FAR (false alarm rate) : Rate of the number of false alarms over all detected change-points 26 Employ three comparison metrics

27. Experimental Results – artificial data – 27 Our alg. with NML was able to detect true change- points and identify the true # of clusters with higher probability than AIC and BIC Average Number of clusters Over Time AIC:Akaike’s information criteria [Akaike1974] BIC:Bayesian information criteria [Shwarz 1978] RNMLAICBIC AR0.9030.1030.135 IR0.3800.0050.020 FAR0.2600.0200.718

28. Comparison w. r. t. KL-divergence Evaluated change detection accuracies by varying the Kullback-Leibler divergence (KLD) between the distributions before and after the change points 28 The larger the KLD between GMMs before and after the change-point was, the more accurately it was detected in terms of IR (identification rate).

29. Experimental Results – vs SW Alg. – SW algorithm : Hypothesis testing whether clusters are identical or not, then make splitting, merging, etc. [Song and Wang, 2005] 29 The sequential DMS with RNML significantly outperformed SW-alg. ARIRFAR Proposed0.9880.9500.050 SW-RNML0.3690.3000.503 SW-BIC0.0190.0000.841 Data : size/time = 512

30. Experimental Results – market analysis – 30 Data set provided by MACROMILL, Inc. Clustering customers to detect their structure changes Our alg. detected clustering changes that corresponded to the year’s ending demand Beer 1Beer 2... User 1350700... User 21050350... Beer 1Beer 2... User 1350700... User 21050350... Beer 1Beer 2... User 1350700... User 21050350... 14 kinds of beer 3185 users 78 days

31. The cluster change in change-point : 1/1,2 31 Many of customers changed their patterns to purchase Beer-A and Third-Beer at the year’s end

32. Conclusion Proposed the sequential DMS algorithm to address clustering change detection issue. Key ideas : Sequential dynamic model selection based on MDL principle The use of the NML code-length as criteria and its efficient computation It is able to detect cluster changes significantly more accurately than AIC/BIC based methods and the existing statistical-test based method in artificial data Tracking changes of group structures leads to the understanding changes of market structures 32

33. Why is NML ? 33 The shortest code-length in the sense of Shtarkov’s minimax criterion [Shtarkov, 1987] Minimum is attained by Ｑ＝ NML distribution Maximum Likelihood Estimator For a given class :

34. Restrict the range of data 34 Restrict the range of data for Shtarkov’s minimax criterion [Shtarkov, 1987] For a given class : Restrict the range of data. We change the Shtarkov’s minimax criterion itself

35. Comparison with non-parametric Bayes Sequential Dynamic Model Selection works better than non-parametric Bayes (Infinite HMM, etc.) [Comparison of Dynamic Model Selection with Infinite HMM for Statistical Model Change Detection Sakurai and Yamanishi, to appear in ITW 2012] 35