Large-scale retrospective event detection from tweets through a DBSCAN-like algorithm in Apache Spark

Messages posted on Location-Based Social Networks (LBSNs) such as Twitter have been reporting everything from daily life stories to the latest local and global news. Monitoring and analyzing this rich and continuous user-generated content can yield unprecedentedly valuable information, enabling user...

ver descrição completa

Detalhes bibliográficos
Autor: Pericacho Sánchez, Gonzalo
Formato: tesis de maestría
Fecha de publicación:2016
País:España
Recursos:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/106430
Acesso em linha:https://hdl.handle.net/2117/106430
Access Level:acceso abierto
Palavra-chave:Parallel algorithms
Big data
Event detection
data clustering
DBSCAN
MapReduce
Apache Spark
Twitter
Algorismes paral·lels
Macrodades
Àrees temàtiques de la UPC::Informàtica
Descrição
Resumo:Messages posted on Location-Based Social Networks (LBSNs) such as Twitter have been reporting everything from daily life stories to the latest local and global news. Monitoring and analyzing this rich and continuous user-generated content can yield unprecedentedly valuable information, enabling users and organizations to acquire priceless knowledge of the occurrences and events. As the size of user-data generated content is extremely large nowadays, parallel processing of complex data analysis becomes essential. In this context, we propose OctreeDBSCAN 3D, an event discovery technique based on a scalable and distributed implementation of the popular density-based clustering algorithm called DBSCAN. First, we evaluate the performance of the proposed algorithm with respect to previous works, achieving speedups of 30 in the improved phases, and second we evaluate the correctness of the clustering performed in real collected data, from which some tagged events were detected. In addition, the tweeting activity varies a lot within geographically large regions and temporarily long periods, limiting the use of global parameters needed for DSCAN-like algorithms. With that in mind, we also present a novel density density-aware MapReduce scheme that partitions tweet data as per its spatial and temporal features and tailors local DBSCAN parameters to local tweet densities. The results from evaluating this scheme pointed out to the benefits of our proposal against other state-of-the-art techniques in terms of speed-up and detection accuracy.