Detecting evolutionary stages of events on social media: A graph-kernel-based approach

Highlights

• We propose a new KPIG graph to represent events on social media.

• We propose a graph-kernel method to detect events’ evolutionary stages.

• We report experimental results on a real dataset and present a case study.

Abstract

Detecting the evolutionary stages of social media events such as Twitter and Sina Weibo is beneficial for enterprises and governments to take necessary actions before emergent phenomena become uncontrollable. Prior work on event extraction from microblogs mostly focused on extracting event summary. However, the evolutionary stages of an event can provide more details about the evolution of the event, which are more critical to predict the future trend of the event. On the other hand, many events have a lifecycle-like evolutionary property, i.e., beginning, developing, climax, descending, and disappearing. Such a feature is helpful to detect the evolutionary stages of events. Thus, we propose a graph-based approach to represent and extract the evolutionary stages of events from microblogs based on this consideration. The contributions of this study are threefold. First, differing from existing methods that use a keyword set or a microblog set to represent an event, we propose a Keyword Popularity Information Graph (KPIG) to represent the keywords and the statistical information of events using a graph. With this mechanism, we can capture both the textual information and the statistical information of an event. Second, based on the KPIG graph, we present a graph-kernel-based approach to measure the similarity among events. Third, we conduct extensive experiments on a real dataset and compare our proposal with several competitor algorithms. The results show that our approach outperforms other competitors in terms of various metrics.

Introduction

Social media analysis has become a hot research topic in recent years [1]. Social media characteristics, such as timeliness, high user participation, and rapid diffusion, provide new opportunities for the early detection and evolution prediction of events. The evolution process of events usually has a specific lifecycle. However, existing research lacks effective methods to accurately extract the evolutionary lifecycle of social media events, which may also affect the accuracy of evolution prediction for events. Detecting the evolution of events helps predict the future trend of events. It is also useful for sentiment analysis over social media because the public sentiment to an event also has the evolutionary property.

Many events in the real world have evolutionary stages, i.e., from birth to death, which is similar to people’s lifecycle. An event’s lifecycle can be defined as a process, including several stages such as beginning, developing, climax, descending, and disappearing, as shown in Fig. 1(a). In addition, there are also some events involving multiple peaks, as shown in Fig. 1(b). The multi-climax event evolution shows an iterative process. However, when we consider social media, it is not a trial task to extract the evolutionary stages of events from social media. There are two challenges:

Challenge 1. The evolution of an event is hard to be captured only by statistical information such as the number of tweets or comments related to the event. For example, if an earthquake event occurred a few days ago, the number of tweets talking about the earthquake will reach a high value. If this number gradually decreases with time and we only consider the number of tweets related to the event, we may conclude that the event has reached the climax and started to degrade. However, this is not always true in the real world. For instance, an earthquake usually incurs many other issues like rescuing, casualties, and city rebuilding. To this end, the event is still at the developing stage. As a result, we need to consider more information like textual information other than statistical event information.

Challenge 2. It is a challenging issue to represent the statistical and textual information of an event effectively. Previous work mainly used keywords or microblogs to describe an event [2], [3], which ignored the linkage information among events and microblogs. Although some works performed evolutionary analysis on events by arranging and linking events according to the timeline [4], [5], they were not enough to detect the evolutionary stages of events because they could measure the similarity among events effectively.

To address the two challenges in extracting the evolutionary stages of events from microblogs, in this paper, we propose a Keyword Popularity Information Graph (KPIG) to represent events and a graph-kernel [6], [7], [8] based approach to detect the evolutionary stages of events. Notably, we make the following three contributions in this paper:

(1) We propose a KPIG graph to integrate textual, statistical, and linkage information for events. Compared to the previous keyword-set based or microblog-set based models, the KPIG graph can provide more rich information about events.

(2) We propose a graph-kernel-based method that runs on KPIG graphs to detect events’ evolutionary stages. Specially, we employ a shortest-path-based graph kernel to measure the similarity and changes between KPIG graphs.

(3) We report extensive experimental results on a real microblog dataset and present a case study to show the effectiveness of our proposal. The results show that our proposal outperforms existing approaches in terms of various metrics.

The rest of this paper is organized as follows. Section 2 provides an overview of the related work. In Section 3, we introduce the KPIG graph. Section 4 presents the graph-kernel-based method to compute the similarity and changes of KPIG graphs. In Section 5, we report the experimental results on a real dataset. Finally, we conclude our work in Section 6.