Information Technology Reference
In-Depth Information
Where:
-
L
is number of all users similar to
u
a
.
-
ω
(
u
j
,act
t
)
is a weighting factor. If user
u
j
did
act
t
,then
ω
(
u
j
,act
t
)=1
,other-
wise
ω
(
u
j
,act
t
)=0
.
-
Parameters
α
satisfies
0
≤ α ≤
1
. It depends on each particular problem.
4.5
Handling Minority Action
In emergency situations, successful actions are often minority actions, and have a great
value. Therefore, we need a method that can handle not only the majority actions but
also the minority actions. However, frequency-based method such as CF can easily
handle the majority actions, but it is difficult to handle minority actions.
To deal with the minority action, we propose the following approach.
1. Using NLP to extract successful actions in feedbacks (tweets) from users. For ex-
ample, if the users said that “it is a
good decision
when staying at company”, then
we can consider “stay at company” as a successful action.
2. Predicting probability of
u
a
who did a successful action, based on the following
idea.
-
Probability of
u
a
who did a successful action is proportional to percentage of
successful actions.
-
The degree of success of an action is proportional to number of good feedbacks
from the users.
Therefore, we can calculate probability of
u
a
whodidasuccessfulactionbyEqua-
tion 5.
DidSuccess
u
a
→act
t
=
f
(
u
a
)
∗
Success
(
act
t
)
(5)
Where:
f
(
u
a
)=
number of successful actions
number of actions
number of good feedbacks about act
t
total number of good feedbacks
Finally, the formula to predict missing action will be complemented as Equation 6.
Success
(
act
t
)=
P
u
a
→act
t
=
DidSuccess
u
a
→act
t
+
P
u
a
→act
t
(6)
5
Evaluation
In this section, we first evaluate our activity extraction approach. Secondly, we use
SPARQL (SPARQL Protocol and RDF Query Language) [18] to evaluate our timeline
action network. Then, we evaluate our proposed approach which predicts missing ac-
tivities. Finally, we discuss the usefulness of the action network.
We collected 416,463 tweets which related to the massive Tohoku earthquake (Ap-
pendix 2). And then, to create data-set for the evaluations, we selected tweets which
were posted by users in Tokyo from 2011/03/11 till 2011/03/12.
