Information Technology Reference
In-Depth Information
Illustration of discovered topics by mmTIM. Reference [
26
]
c
Fig. 4.6
2012 Association for
Computing Machinery, Inc. Reprinted by permission
dramatically during the first several iterations and converges to a stable level within
20 iterations. Since larger topic number requires more computation cost and captures
weak semantics, we prefer the smallest
T
that yields small perplexity and fast con-
vergence. It is clear that the perplexity decreases much slower when
T
20 and we
choose the desired topic number
T
to be 20 for mmTIM in the following experiments.
>
4.5.2.2 Illustration of Discovered Topics
In order to interpret the derived latent topic space, we visualize two of the discovered
20 topics by providing five top-ranked tag words and the most five related images. As
represented in Fig.
4.6
, the tagwords are sorted by their probability of being generated
from the corresponding topic
p
, while images are sorted by counting the topic
indicator variables of visual descriptors and tag words
p
(
w
i
|
Z
t
)
(
Z
t
|
w
d
,
v
d
)
:
n
d
i
)
+
n
d
i
z
d
,
i
z
d
,
i
1
I
(
=
t
1
I
(
=
t
)
=
=
p
(
Z
t
|
w
d
,
v
d
)
=
,
n
d
+
n
d
where
z
d
,
i
and
z
d
,
i
are the topic assignments for the
i
th tag word and visual descriptor
for image
d
, and
I
is indicator function returning 1 is it is true and 0 otherwise.
By providing a combination of representative words and image, it becomes very
easy to interpret the domain knowledge associated with each topic. We can see that,
by considering both textual tag words and visual image content, the discovered latent
topics show high consistency between semantic concepts and visual themes.
(
ยท
)
4.5.2.3 Qualitative Case Study
We demonstrate the effectiveness of mmTIM on topic-sensitive influencer identifi-
cation of sampled users. To get a overview of the derived topic-sensitive influence,
we first visualize the influence network for two test users on topic #2 and #13, in
Fig.
4.7
. The width of the arrow is proportional to the influence strength. From a
global view, we can see the peer-to-peer influence strength is much different from
topic to topic. To look into more detail from a local view, we check the widest arrow
in the influence network to investigate the two test users and the influencers who
