IEEE Computational Intelligence Magazine - August 2019 - 36

Whenever the text stream starts to change and deviate
from its current trending topics, the proposed WDHG
method quickly detects the emerging event due to its
sensitivity towards the dynamic nature of the text stream.
Second, we compare the performance
of the winner approach (i.e., aggregated
centrality) as the event detection method,
and perform a comparative analysis with
the following state-of-the-art methods:
❏ Probabilistic Model - Latent Dirichlet
Allocation (LDA) [19]
❏ Clustering - Document-pivot (Doc-p)
[20], BN-gram [12]
❏ Frequent Pattern Mining - Soft Frequent Pattern Mining (SFPM) [12]
❏ Matrix Factorization - SVD-KMean
[21], SNMF-Orig, SNMF-KL [18]

Topic-Recall

FA Cup

❏ Exemplar-Based - Exemplar [15]
❏ Graph-based - Graph-based Feature-

pivot (GFeat-p) [22]
For the FA Cup dataset, Figure 7(left)
shows the results for T-Recall at K =
{2, 4, 6, f, 20}. For all the approaches,
the results for the FA Cup are the best
among all the three datasets due to short
duration and high popularity of the football match. Users posting tweets for such
events are focused and consistent. Therefore, topics that appeared in the data are
less diverse and easier to detect than in

Super Tuesday

US Election

0.8
0.6
0.4
0.2
2
4
6
8
10
12
14
16
18
20
2
10
20
30
40
50
60
70
80
90
100
2
10
20
30
40
50
60
70
80
90
100

0
Top-K
LDA
SVD-Kmean

Doc-P
SNMF-Orig

Gfeat-P
SNMF-KL

SFPM
Exemplar

BNGram
WDHG

FIGURE 7 Comparison with baseline methods for FA Cup, Super Tuesday and the US Election
datasets.

TABLE 2 Comparison with baseline methods for Keyword-Precision@2 for FA Cup,
Super Tuesday, and the US Election datasets. Best results for each dataset are in
bold face.

Super Tuesday and the US Election datasets. For smaller values of K, WDHG
does not perform well; however, it quickly gains the maximum T-Recall and
detects all the ground truth topics at K =
20 as shown in Figure 7(left).
Conversely, the Super Tuesday and
the US Election datasets have more
diversity in their text streams due to
greater temporal coverage compared to
the FA Cup. The fundamental characteristic of the proposed method is to sense
change in the text stream and to extract
coherent topics from the word-graph
(see Section 2). Therefore, the proposed
method outperforms all other baselines
at K 2 50 for the Super Tuesday dataset
as shown in Figure 7(middle).
The US Election is one of the largest
datasets in terms of temporal coverage
which spans over 36 hours including
sixty 64 topics. The proposed WDHG
method significantly outperformed all
the baseline methods at K 2 2 as shown
in Figure 7(right).
Table 2 shows the results of K-Precision at K = 2. The proposed WDHG
method combines the scores of the displaced temporal frequency and the topic
centrality in the word-graph; therefore it
detects even-related keywords with high
precision for all three datasets. The proposed WDHG method is an effective
approach and performed consistently
better compared to existing methods in
terms of performance and accuracy.
We observe that whenever the text
stream starts to change and deviate from
its current trending topics, the proposed
WDHG method quickly detects the
emerging event due to its sensitivity
towards the dynamic nature of the text
stream. It also appears that user participation increases when an event occurs.
The user participation feature is not
used in the proposed WDHG method;
however, it can be useful to improve the
detection model further. We are planning to use it in future studies.

METHOD

FA CUP

SUPER TUESDAY

US ELECTION

LDA

0.164

0.000

0.165

Doc-P

0.337

0.511

0.401

Gfeat-P

0.000

0.375

SFPM

0.233

0.471

0.241

BNGram

0.299

0.628

0.405

SVD-Kmean

0.242

0.367

0.300

3.5. Execution Time Analysis

SNMF-Orig

0.330

0.241

SNMF-KL

0.242

0.164

Exemplar

0.300

0.485

0.391

WDHG

0.545

0.750

0.423

Twitter generates a great number of
micro-documents within a short interval, and frequency of publishing microdocuments increases rapidly when a

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | AUGUST 2019

IEEE Computational Intelligence Magazine - August 2019

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