1. ADAPTIVE EVENT DETECTION USING TIME-VARYING POISSON PROCESSES Kdd06 University of California, Irvine

2. ABSTRACT Time-series of count data  aggregated behavior of individual person  periodic  bursty periods of unusual behavior In this paper  statistical estimation techniques  time-varying Poisson process model  unsupervised learning Two data sets with ground truth  freeway traffic data  building access data  performs better than a non-probabilistic, threshold-based technique

3. CONTENT Introduction Related work Data set characteristics A baseline model and its limitations Probabilistic modeling Learning and inference Adaptive event detection Estimating event attendance Conclusion

4. INTRODUCTION

5. Focus on  time-series data where time is discrete and N(t) is a measurement of the number of individuals or objects recorded over the time-interval [t-1, t].

7. DEFINITION OF EVENT Events  Sustained (bursty) periods of anomalous behavior  sometimes refer to individual measurements  Here, a large-scale activity that is unusual relative to normal patterns  such as a large meeting in a building, a malicious attack on a Web server, or a traffic accident on a freeway. Chicken and egg problem  requires some knowledge of what constitutes normal behavior  historical data consists of both normal and anomalous (event) data mixed together.

8. Goal  define a model of uncertainty (how unusual is the measurement?), and additionally incorporate a notion of event persistence.  learn a model that reflects the bimodal nature of such data, namely a combination of the normal traffic patterns to which is occasionally added additional counts caused by aperiodic events.

9. RELATED WORK

10. Techniques  Markov model  Likelihood-based method  A combination of Poisson models and Bayesian estimation methods  Infinite automaton Common goal  Detect novel and unusual data points or segments in time-series

11. DATA SET CHARACTERISTICS

12. BUILDING DATA

13. FREEWAY TRAFFIC DATA

14. Holiday data should be removed before modeling, because they involve relatively different behavior.

15. A BASELINE MODEL AND ITS LIMITATIONS

16. Threshold test based on a Poisson model  estimate the Poisson rate λ of a particular time and day by averaging the observed counts on similar days at the same time  The max likelihood estimate  and λ < N

18. Limitations  Is adequate when events cause a large increase in count data  Fail when facing the chicken and egg problem  Thresholds and the false alarms

19. PROBABILISTIC MODELING

20. Model N(t)  Normal behavior: N 0 (t)  Event caused: N e (t) 

21. MODELING PERIODIC COUNT DATA Poisson distribution  λ (t)  d(t)  Indicates the weekday on which time t falls h(t)  Indicates the interval in which time t falls δ and η 

22. MODELING PERIODIC COUNT DATA The effect of δ d(t)

23. MODELING PERIODIC COUNT DATA The effect of η d(t),h(t)

24. MODELING PERIODIC COUNT DATA

25. MODELING RARE, PERSISTENT EVENTS Use binary process z(t) to indicate the presence if an event  Transition probability matrix   Length of period between events is with expected value 1 / z 0  Length of each event is with expected value 1 / z 1 z 0 and z 1 priors 

26. N E (t)  γ (t) is independent at each time t 

27. Markov-modulated Poisson model

28. LEARNING AND INFERENCE

29. MCMC  Markov chain Monte Carlo methods  Monte Carlo 方法的基本思想是 ：为 了求解某个 问题， 建立一个恰当的概率模型 或随机 过 程 ， 使得其参量 （ 如事件的概率 、 随机 变 量的数学期望等 ） 等于所求 问题 的解 ， 然后 对 模型或 过 程 进 行反复多次的随机抽 样试验， 并 对结 果 进 行 统 计 分析 ， 最后 计 算所求参量 ， 得到 问题 的近似解 。 Hidden variables  {z(t), N 0 (t), N E (t)}

30. SAMPLING THE HIDDEN VARIABLES GIVEN PARAMETERS Likelihood functions  Sample  If z(t) = 0  N 0 (t) = N(t) If z(t) = 1 

31. SAMPLING THE PARAMETERS GIVEN THE COMPLETE DATA Integral number of weeks  T = 7 * D * W The complete data likelihood  In this case, only involve λ 0, δ and η  Sufficient statistics of the data

32. Posterior distributions 

33. ADAPTIVE EVENT DETECTION

38. ESTIMATING EVENT ATTENDANCE

40. CONCLUSION

41. Described a framework for building a probabilistic model of time- varying counting processes Observe a superposition of both time-varying but regular(periodic) and aperiodic processes Applied this model to two different time series of counts both over several months Described how the parameters of the model may be estimated using MCMC sampling