IEEE Computational Intelligence Magazine - August 2019 - 34

2.4. Trending Topics

Multiple candidate WDHGs within a
sliding window can have duplicate
words. We select only unique words that
have the highest weights among candidate graphs to extract the event-related
topics without compromising perfor-

mance. The weight for each word is calculated by fusing degree centrality and
displaced temporal frequency scores. The
example shown in Figure 4 further
elaborates the process of merging multiple candidate WDHGs, where WDHG ci
is the i-th candidate graph in the sliding

WDHGs at Three Subsequent Time Intervals
Pre-Event at ti -1

Event at ti

Post-Event at ti +1

Weak

Strong

Weak

WDHGs
Visualization

window under process. In the sliding
window kTt, a ranked list of topics is
extracted from all candidate WDHGs
and classified as "Strong". Topic ranks are
calculated using (12).
Rank(v }k ) =TC (v }k ) # j (v k}) .

(12)

Figure 5 shows the visualization of
three consecutive WDHGs with their
class labels and the top ten keywords
from an event (i.e., Goal) in the FA Cup
dataset. The event detection model is
described in Algorithm 2. It consists of
feature extraction, event candidate
detection, and finally extraction of an
event-related ranked list of topics.
3. Experiments and Evaluation

Class Label

Top-10
Event-Related
Keywords

watch
win

amp
liverpool
goal

chelsea
score

game

ramir
1-0

FIGURE 5 Visualization of three WDHGs G hi -1, G hi , and G hi +1 at time t i -1, t i, and t i +1, respectively using the FA Cup dataset. A significant event "Goal" occurred at interval t i . The nodes
and edges with positive and negative weights are shown in red and green, respectively. A large
portion of the graph is affected when an event occurs. It shows that WDHG is hyper-sensitive
to detect newly emerging topics.

Algorithm 2 Event Detection Algorithm.

input: Gh (kTt) - Set of WDHGs within the sliding window kTt
output: L - List of ranked topics
foreach G hi ! Gh (kTt) do
calculate growth factor GFi using (1)
calculate aggregated centrality AC i using (5)
calculate heartbeat score HB i using (7)
end
calculate i (kTt) using (11) for Gh (kTt)
foreach WDHG G hi ! Gh (kTt) do
assign binary class label C to G hi for the corresponding heartbeat score HB i using (8)
if C = "Strong" then
assign weight to each keyword in G hi using (12)
merge keywords from G hi into the topic list L
remove duplicates from L while keeping keywords that have maximum weight

1
2
3
4
5
6
7
8
9
10
11
12
13
end
14 end
15 sort L based on topic ranking

34

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | AUGUST 2019

We propose three event detection approaches. These approaches are based on
the measures growth factor and topic
centrality and their combination (see
Subsection 2.3). To test and evaluate
these three approaches, we first compare
their performances with each other; then
the best approach is selected and compared with nine state-of-the-art methods
(see Subsection 3.4). We perform experiments on three benchmarks: FA Cup,
Super Tuesday and the US Election datasets. An existing evaluation framework is
used for comparative analysis [12].
3.1. Parameters

Three parameters (i.e., Tt, t i, and ~) are
used for optimization. Tt is the temporal
coverage of the sliding window that
contains a batch of data to be processed
independently. As per ground-truth, the
temporal coverage Tt of sliding window
is set to one minute, one hour, and ten
minutes for FA Cup, Super Tuesday, and
the US Election datasets, respectively.
The small size of the micro-document in
the Twitter stream did not yield useful
information; therefore, micro-documents
are accumulated together that appear
within a specific time span to form a
super-document. The parameter t i, a
temporal coverage of super-documents,
is set to one minute, ten minutes, and
one minute for the FA Cup, Super Tuesday, and the US Election datasets, respectively. ~ is the adjustment parameter that
IEEE Computational Intelligence Magazine - August 2019

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