Navegando por Palavras-chave "Detecção De Eventos"

Agora exibindo 1 - 1 de 1
  • Nenhuma Miniatura disponível
    Item
    Somente Metadadados
    A Comparative Study On Regression Approaches For Event Detection In Instagram
    (Universidade Federal de São Paulo (UNIFESP), 2017-11-30) Santos, Elder Donizetti Dos [UNIFESP]; Faria, Fabio Augusto [UNIFESP]; Universidade Federal de São Paulo (UNIFESP)
    With the advancement of the use of web-based applications and mobile device technologies, in particular, online social networks, many approaches have been proposed in the literature using it as the source of information. Online social networks like Instagram have more than 700 million users who, together, create over 300 million new content every day. All of this data can be used, for instance, to detect real-world events. Such events can be defined as a car accident, a natural disaster, a riot, a political affair, among others. In order to do that, algorithms need to manage massive, rapidly changing and fast arriving data streams made of text, images, and videos. It also involves challenges such as the lack of a labeled database to analyze the effectiveness of applied techniques that can be reused by other researchers and the need for an approach that adapts to the constant changes in the flow of information. However, existing approaches are often either limited or not suitable for new data sources like Instagram. In this sense, this work provides contributions in the area of event detection for online social networks. As a first contribution a review on how the task of event detection has been approached by researchers since its inception in the 1990’s is presented. The second contribution is an introduction to the behavior and volume characteristics of Instagram posts modeled as time series. Then, a comparative study of different regression techniques for time series prediction is conducted by applying a preprocessing step and algorithms such as Support Vector Regression (SVR), Multilayer Perceptron (MLP), Autoregressive Integrated Moving Averages (ARIMA), Classification and Regression Trees (CART) and K-Nearest Neighbors (KNN). As a result, it is demonstrated how a simple yet efficient approach can be used to detect events in social networks. Trying to overcome some of the challenges mentioned, as a third contribution, a semi-supervised learning approach is proposed using time series correlations. Experimental studies have shown that time series from different sub-regions with similar characteristics can be used to generalize knowledge and predict the occurrence of an event. Moreover, it is demonstrated that the proposed approach is a good alternative to the Gaussian Process Regression (GPR) used in the literature since the approach based on time series correlations provides good results using much less computing resources than GPR. In addition to the main contributions cited, the entire dataset used in this thesis with more than 180 thousand manually labeled Instagram posts is publicly available.