Open Collections

Poisson Process Infinite Relational Model : a Bayesian nonparametric model for transactional data

University of British Columbia

Transactional data consists of instantaneously occurring observations made on ordered pairs of entities. It can be represented as a network---or more specifically, a directed multigraph---with edges possessing unique timestamps. This thesis explores a Bayesian nonparametric model for discovering latent class-structure in transactional data. Moreover, by pooling information within clusters of entities, it can be used to infer the underlying dynamics of the time-series data. Blundell, Beck, and Heller (2012) originally proposed this model, calling it the Poisson Process Infinite Relational Model; however, this thesis derives and elaborates on the necessary procedures to implement a fully Bayesian approximate inference scheme. Additionally, experimental results are used to validate the computational correctness of the inference algorithm. Further experiments on synthetic data evaluate the model's clustering performance and assess predictive ability. Real data from historical records of militarized disputes between nations test the model's capacity to learn varying degrees of structured relationships.

Text

2016-08-22

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/58921

Statistics

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/