Supporting real-time mobility services with scalable flock pattern mining

• Autor(a) principal: LACERDA, Thiago de Barros
• Data de Publicação: 2016
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Institucional da UFPE
• dARK ID: ark:/64986/001300000rjmk
• Texto Completo: https://repositorio.ufpe.br/handle/123456789/18700
• topic: as-chave: CidadesInteligentes. MineraçãodePadrões. DadosEspaço-temporais. Padrão de Flock.; SmartCities. PatternMining. Spatio-temporaldata. FlockPattern.

Pattern mining in spatio-temporal datasets is a really relevant subject in the academia and the industry nowadays, due to its wide applicability in helping to solve real-world problems. Many of them can be found in the context of Smart Cities, like Traffic Management, Surveillance and Security and City Planning, to name a few. Among the various spatio-temporal patterns that one can extract from a spatio-temporal dataset, the flock pattern is one that has gained a lot of attention, because of its intrinsic relation with the aforementioned problems. A lot of work has been done in the academia, in order to provide algorithms able to identify the flock pattern. However, none of them could perform that task efficiently nor be able to scale well when a large dataset was the analysis target. Additionally, we found that there was no system architecture proposal that could be simple and modular enough to be used in that spatio-temporal pattern detection problem. Given that context, this dissertation proposes a modular system archicture designed to help solving flock pattern mining problems and possibly be reused to other spatio-temporal mining experiments. We then use such architecture as the infrastructure to implement an efficient flock detection algorithm, aiming at achieving considerable gains in execution time without compromising accuracy, thus targeting real-time deployment and on-line processing in Smart Cities. Last, but not least, we remodel our algorithm in order to take advantage of multi-core architectures present in modern computers. Our results indicate that our proposal outperforms the current state-of-the-art techniques, by achieving 99% CPU time improvement. Moreover, with our multi-thread model, we were able to reduce the processing time of our proposed algorithm by 96% in some cases. We prove the efficiency of our solution by performing evaluation with both real and synthetic large datasets.