International Journal of Hybrid Information Technology
Vol.8, No.9 (2015)
188 Copyright ⓒ 2015 SERSC
represents the total usage volume that u
i
used s
j
under C
l
.
is the mean
o f t h e
usage volume that u
i
used s
j
under C
l
and it is represented as follows:
llq
lqjilji
Cc
csu
l
Csu
Vol
n
Vol
,,,,
1
(3)
Where, n
l
represents the number of context instances that C
l
contains. n
l
will change
when the different context partition methods are used. Introduced by the case of the time
context, if it is divided into morning, afternoon, evening, night, n
l
=4; if it is divided into
hour, n
l
=24.
The larger the volatility is, the greater the impact that context towards mobile user
preference is. Introduced for the case of time context and location context, the voice call
duration of some mobile user under the given context is as follows: {morning, afternoon,
evening, night}={20,20,60,0}, the volatility obtained by Equation (2) is 0.7; {at home, at
work}={52,48}, the volatility obtained by Equation (2) is 0.04. We can know that the
voice call affected by the time context is greater than that by the location context in the
above example. In the paper, we need to set the threshold of the volatility. If the obtained
volatility is greater than the setting threshold, we judge the mobile user preference is
affected by the given context; else, we judge the mobile user preference isn’t affected by
the given context.
The principle that sets the threshold of the volatility is as follows. When the threshold
is relatively small, it’s loose for the volatility, so there are many contexts that are judged
to impact the mobile user preference. There is a Cartesian Product proportional
relationship between the number of contexts and the number of the mobile user
preferences [13]. Therefore, there are many mobile user preferences and the running time
is longer, but the accuracy of learned mobile user preferences is higher. When the
threshold is relatively great, the contexts that are judged to impact the mobile user
preference are few. There are few mobile user preferences and the running time is shorter,
but the accuracy of learned mobile user preferences is incomplete. In the above example,
when the threshold of the volatility is set 0.1, the voice call is affected only by time
context and the number of the context instances is 4; when the threshold of the volatility
is set 0.01, the voice call is affected by the time context and the location context, and the
number of the context instances is 8(4*2). Therefore, it needs to compromise between the
accuracy and the running time when setting the threshold of the volatility.
We can calculate the impact that context towards the mobile user preference according
to Equation (4):
j
s
jim
jil
jil
n
m
suC
suC
suC
vol
vol
w
1
,,
,,
,,
(4)
Where,
represents the number of contexts that impacted u
i
when he used s
j
.
I n t e g r a t i n g t h e
obtained weight of context and the usage volume that mobile user used mobile network
service under given context instance, the weight of context instance can be calculated by
Equation (5):
lji
lqji
jiljilq
Csu
csu
suCsuc
Vol
Vol
w
,,
,,
,,,,
*
(5)