IEEE Computational Intelligence Magazine - August 2019 - 30

The social media stream such as Twitter provides
real-time information daily about worldwide events
and generates a huge amount of data, in terms of
content and diversity.
dynamic and non-stationary data. They
focus on bursty features that are frequent,
co-occurring, and biased toward highly
weighted patterns. These methods are
well-placed to detect event-related topics
with high frequency but ignore the dominant nature of burstiness due to which
smaller events in the data are overlooked.

to describe an event. Thus, sophisticated
techniques are required to detect events
on Twitter.
Recent work on event detection
heavily relies on bursty features [4], [5].
However, bursty features may not be able
to detect events that are less popular. In
this study, we develop a robust and sensitive graph-based approach. The proposed
approach identifies significant changes in a
text stream to detect events.This approach
uses displaced temporal frequency, which
is a change in frequency with respect to
time, to characterize the event-related
keywords and their dynamic relations.

1.2. Contributions

In this study, we propose an approach
Weighted Dynamic Heartbeat Graph
(WDHG) to detect events. Once an
event is detected, WDHG approach suppresses the bursty keywords at subsequent time intervals. This enables other
related information to be more visible
and helps in capturing new and emerging events. The proposed approach is
change-sensitive and detects event-related topics efficiently. It repeatedly captures the change-patterns in the time
series data of Twitter streams and highlights key occurrences.
The key contributions of our study
are, therefore:
❏ The proposed approach (WDHG) is
change-sensitive, as it subsequently
suppresses high-frequency events
after their first detection.

1.1. Problem

The analysis of Twitter stream, which
contains diverse and noisy content,
requires that a number of challenges are
addressed such as accuracy, efficiency, and
scalability. Micro-documents that are published to report an event have a set of
related keywords that can be used to
detect an event and to identify a relevant
description. To identify significant information patterns, many graph-based methods have been proposed to deal with
real-life data [6]-[9]; however many are
inadequate for the analysis of complex,

Tweets (min)
5 min Moving Average

2,500
2,000
1,500

First Goal

Second Goal

Match
Started

End of Match

Third Goal
Goal
Missed

Half Time

1,000
500

15:30
15:40
15:50
16:00
16:10
16:20
16:30
16:40
16:50
17:00
17:10
17:20
17:30
17:40
17:50
18:00
18:10
18:20
18:30
18:40

FIGURE 1 Tweets traffic at the time of Final Match of Football Association Challenge Cup, 2012
(FA Cup) between 15:30 and 18:40.

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | AUGUST 2019

❏ WDHG is computationally efficient.

It detects events in polynomial time.
❏ Experiments on three real-life bench-

mark datasets: FA Cup, Super Tuesday
and the US Election demonstrate
that WDHG outperforms state-ofthe-art methods.
2. Weighted Dynamic Heartbeat
Graph (WDHG) Approach

Figure 2 shows the workflow of the
proposed approach and steps involved to
transform the Twitter text stream into
temporal graphs for extracting eventrelated topics. The data undergoes several transformations starting from the text
stream as input. Preliminary concepts
and definitions can be found here [10].
Micro-documents in the text stream are
accumulated to generate super-documents. A series of graph inheriting word
co-occurrence relationships from the
micro-documents are generated using a
set of super-documents. Furthermore,
each adjacent pair of graphs is mapped
onto a WDHG (see Subsection 2.2).
Afterward, features are extracted (see
Subsection 2.3). A rule-based classifier
labels the candidate WDHGs for event
representation. To classify WDHGs as
event candidates, we use aggregated
centrality as a key feature. Finally, all the
candidate WDHGs are merged to
extract event-related topics (see Subsection 2.4).
This study is an extension to our previous work [11]. Unlike our previous
work, we construct a weighted graph
structure and use three different approaches
for the event detection model (see Subsection 2.3). We evaluate the performance
of each detection model (see Subsection 3.4).
We extend experiments with a larger
benchmark dataset (i.e., the US Election)
for the evaluation. We also perform a
detailed execution time analysis to determine the efficiency of the proposed approach (see Subsection 3.5).
2.1. Assumptions

For detecting events, the proposed
approach works with the following assumptions:
❏ There is a significant change in the
temporal frequency of words, or new

IEEE Computational Intelligence Magazine - August 2019

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - August 2019