Sensors 2016, 16, 1706 6 of 26
Context describes the information which characterizes the situation of the entity in the ubiquitous
computing. Emerging sensor networks can provide us much abundant context information.
The widely
used context usually reflects the physical environment (e.g., location, time). However, some
contexts can include other types of data, for instance, information about the user (the user’s habits,
bio-physiological conditions, etc.), physical conditions (noise, light, temperature, etc.),
and social
environment (social interaction, co-location with other users, etc.) [
29
]. In this system, we observe the
context of information about the user and physical conditions, and the context can be given as a vector
of different context types [30]:
C = (C
1
, C
2
, ···C
z
) (1)
where C
t
(t ∈ 1,2, . . . ,z) is a context type, such as location, time, temperature, and so on.
The Pearson Correlation Coefficient method is adopted to measure the simility between two
different contexts Sim(x,y). Then, we have:
Sim(x, y) =
n
∑
xy −
∑
x
∑
y
q
n
∑
x
2
−(
∑
x)
2
×
q
n
∑
y
2
−(
∑
y)
2
(2)
where x and y are two different contexts.
Based on Equation (2), we define the number of operations or activities done by the user u on the
device i in the context x as r
u,x,i
:
rel(x, y, i) = k
∑
u∈U
(r
u,i,x
t
−r
i
) ×(r
u,i,y
b
−r
i
)
σ
x
×σ
y
(3)
where k is a coefficient that is used to adjust the sensity of the relavance, and
r
i
is the average number
of operations. U represents the set of all users in the system.
σ
x
,
σ
y
are the standard deviations for
the two contexts. rel(x,y,i) returns the relevance of the two context values in C over all the number of
operations done by users. To get a better and clearer result, we can incorporate this relevance feature
into the similarity calculations. Besides, context information usually involves in privacy problems.
For this reason, in our recommender system, the information data is encrypted.
For most item-based collaborative filtering algorithms the similarity (conditional probability-base
similarity, cosine-based similarity, etc.) between different items can be calculated by analyzing historical
use or purchasing data, which is usually presented as a user-item matrix. Then, a recommendation list
which is sorted by using some interest measure for items will be derived in a Contextual Item-based
Collaborative Filtering (CICF) recommender system [
31
]. The main idea of collaborative filtering
is to predict the items that people will buy or prefer according to what they liked or bought in the
past [32,33]. The process of the contextual item-based collaborative filter algorithm has three steps:
1. Disposal of the historical data about the users and items and building of the user-item matrix.
2.
Calculation of the similarity between each pairs of items, and building of an item-item
similarity matrix.
3.
Calculation of the current user’s location-aware taste for each item by the item-item similarity
matrix and the user’s historical record, and choosing the most interesting items to generate
the recommendations.
Different from the case of desktop calculation, there is more abundant context information in the
wireless sensor environment, and the recommendation systems that can achieve better performance
benefits from the context, for example, location or time [
34
]. CICF is also based on the idea of collective
intelligence, and brings contextual information into the similarity of items and collaborative filtering
model. The key of CICF is that calculates preference similarities of items in the context condition.
Chen proposed a context-aware collaborative filtering recommender system that integrates
contextual information of items and user-context information into the collaborative filtering,