Information Technology Reference
In-Depth Information
We can see that mmTIM shows its capability in identifying the most influential
contact users on topic level. The identified influencers have high #follower and show
strong expertise on the corresponding topics. For example, user “95386698@N00”
has major interest in topic #2 from his/her topic distribution. From the uploaded
exemplary images and tag cloud, we can see he/she has conducted extensive activities
related to topic #2. Combining the large #follower and the decent design and fashion
images uploaded by user “26324110@N00”, we may roughly conclude that he/she
is an expert on fashion topics, i.e., topic #13.
4.5.2.4 Quantitative Evaluation
Now we conduct quantitative evaluation of topic-sensitive influencer mining. Recall
that the goal is to identify contact users with themost influence to the target user on the
specific topic. Observing that many of the images to which user add
Favorite
marks
are uploaded by his/her contact users, we leverage user's
Favorite
images to generate
ground-truth information for the evaluation. Specifically, for each
Favorite
image
d
of test user
U
i
, we calculate its topic proportion and assume the image belongs to the
dominant topic, which is denoted as
z
d
. On each topic, the contact user
U
j
owning
the top #images
Favorited
by
U
i
is considered most influential to
U
i
on this topic.
Formally, the most influential contact user to user
U
i
on the
t
th topic is defined as:
I
t
U
i
=
argmax
I
(
z
d
=
t
,
u
d
=
U
j
)
U
j
∈
C
U
i
;
d
∈
F
U
i
where
u
d
is uploader of image
d
,
F
U
i
is
Favorite
image set of
U
i
.
We consider the following two topic-sensitive influence modeling methods for
comparison:
•
Topical Affinity Propagation (TAP [
30
]): a method learning topic space and topic-
sensitive influences separately by inputting the nodes' topic mixtures;
•
Mining Topic-level Influence on Heterogeneous networks (mTIH [
17
]): a proba-
bilistic model exploiting the heterogeneous link and textual content information,
which targets at text-based citation networks.
Note that the topic space of TAP is pre-extracted by running a standard LDA on the
user-annotation corpus with each user's annotation set as one document and specific
tag as word, and the citation link between users in [
17
,
30
] is replaced by the contact
relation in our implementation.
We utilize top-k accuracy as the evaluation metric. For each test user
U
i
,we
rank the contact users by their influence values on the
t
th topic. Denoting the rank
of ground-truth influencer
I
t
U
i
I
t
U
i
)
C
U
i
ˀ(
in contact list
as
, the top-k accuracy is
calculated as
t
,
U
i
I
t
U
i
)
≤
I
(ˀ(
k
)
Accuracy
(
k
)
=
(4.18)
T
·|
U
|
