Nonparametric approaches for discovering triggering events from spatio-temporal patterns

---

Tezin Türü: Doktora

Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Enformatik Enstitüsü, Bilişim Sistemleri Anabilim Dalı, Türkiye

Tezin Onay Tarihi: 2014

Öğrenci: BERNA BAKIR BATU

Danışman: TUĞBA TAŞKAYA TEMİZEL

Özet:

Temporal or spatio-temporal sequential pattern discovery is a well-recognized important problem in many domains such as seismology, criminology and finance. The majority of the current approaches are based on candidate generation which necessitates parameter tuning such as definition of a neighborhood, an interest measure and a threshold value to evaluate candidates. However, their performance is limited as the success of these methods relies heavily on parameter settings. In this thesis, two sequential pattern mining algorithms are developed for the multi-type spatio-temporal point patterns based on the nonparametric stochastic declustering methodology. The algorithms use multivariate conditional intensity model to define triggering relations within and among the event types and employs the estimated model to extract significant triggering patterns. They initially estimate pairwise triggering probabilities of all instances according to the multivariate Hawkes model, and then generate candidate patterns by using a rank selection method. Since a pair of instances is associated with a triggering probability, the proposed approaches also allow user to evaluate the significance of the pairwise pattern of any event type.The proposed methods are tested with synthetic data sets exhibiting different characteristics. The method gives good results that are comparable with the methods based on candidate generation in the literature. It is observed that the discretization of the density function based on the significant interaction ranges obtained by Diggle D-function maximizes the triggering probabilities of the patterns that exist at similar scales. The method is tested with real data to estimate the effects of the speed bumps on the number of accidents reported in METU Campus.