Page 148 - ITU KALEIDOSCOPE, ATLANTA 2019

2019 ITU Kaleidoscope Academic Conference

reasoning and the ability to acquire new knowledge. However, similarity between keyword x i and y j .
due to the complex care demands of various geriatric
diseases, the traditional knowledge expression methods are x 1 y 1 x 1 y 2 · · · x 1 y m
not suitable for care demands decomposition, such as logical © x 2 y 2 · · · x 2 y m ®
ª
­ x 2 y 1
representation, semantic network, rule-based system and so S XY = ­ . . . . . . . . ® (3)
­
®
on. Therefore, we designed the expert knowledge forest to ­ . . · · · . . ®
combine care demands and expert knowledge based on the « x n y 1 x n y 2 x n y m ¬
decision tree structure and the frame knowledge expression. According to [20], the semantic similarity between keywords
In the expert knowledge forest, every frame consists of five x i and y j is calculated based on path length, depth and local
expert knowledge components: (1) name, a unique name density as shown in Equation (4), where l represents the path
that can be any constant; (2) slot, a combination of expert length and h represent the path depth. Simultaneously, α
knowledge and care demand; (3) slot value, the attribute is a constant and β is a smoothing factor where β > 0. In
value, which can be 0 or 1; (4) relation, the knowledge this paper, we set α = 0.2 and β = 0.6 to generate optimal
associations between frames; (5) slot constraints, related semantic similarity, as details are shown in [20].
constraints contributed to the corresponding slot value.
e βh − e −βh
−αl
s x i , y j = e βh −βh (4)
4.2 Smart-desire model e + e
In order to normalize the fuzzy similarity matrix, we
Due to the limitations of family space and expenditure, compress it into one dimension by taking the maximum value
designers need to extract suitable smart services to provide for each row of the matrix, and average these maximum values
geriatric care for the elderly in the smart home. In this paper, by Equation (5). However, s_sem (X,Y) only represents the
we proposed a Smart-desire model to extract smart services average semantic similarity between the vector B and every
based on expert knowledge forest and mapping similarity. word in A. Thus, we use Equation (6) to calculate the semantic
As shown in Definition 2, S desire is the output of SDMM, similarity between vector X and vector Y.
and P UED is proposed to decompose care demands(D)
by traveling expert knowledge forest (E) based on user 1 n Õ
s_sem (X,Y) = × max x i y j ; j ∈ [1,m] (5)
information(U). Additionally, in the Smart-desire model, SS n
∗
∗
is a set of all smart services in the service layer, and ϕ (x ,y ) i=1
is the mapping method in SDMM. (s_sem (X,Y) + s_sem (Y, X))
s_sem (|X,Y|) = (6)
2
Definition 2 Smart-Desire is presented to obtain S desire Therefore, the functional similarity between care demand c
based on user information and expert knowledge. and smart service s is calculated by Equation (7), where ϕ 1
and ϕ 2 are weight coefficients of service name and service
∗ ∗
Smart_Desire = {SS, P UED , ϕ (x ,y ),S desire } . description, with ϕ 1 +ϕ 2 =1.
Õ
S desire = ϕ (P UED ,SS) = S i .
f _sim(|c, s|) =ϕ 1 × s_sem (|c_Description, s_Description|)
ϕ 2 × s_sem (|c_Label, s_Label|)
4.3 Mapping similarity calculation method (7)

4.3.1 Functional similarity calculation
4.3.2 Non-functional similarity calculation
In the Definition module of the smart service, the Label and Since the QR of user demands are difficult to accurately
Description describe the function of this service with phrases, represent with a number, we propose a scoring set to
sentences, or concepts. In order to weight the importance of standardize these parameters, as shown in Equation (8).
words in the Description of smart services and care demands,
we employ the famous term frequency-inverse document
frequency (TF-IDF) method. After calculating the TF-IDF score ∈ {Excellent, Good, Medium, Poor, Very Poor} (8)
values of words in Description of smart service s j , we obtain
the vector X and vector Y by ranking these words according Although the scoring set can express the level of QR
to TF-IDF values, as shown in Equation (1) and Equation (2). parameters, it cannot be directly used for calculation.
Therefore, the intuitionistic fuzzy theory is used to quantify
these QR parameters. In this paper, since the scoring set
X = (x 1 , x 2 ,..., x n ) (1)
is defined as 5 levels, the intermediate value is defined as
0.5, and the remaining distribute symmetrically. According
to the fuzzy number of the scoring set in Table 1, we use
Y = (y 1 , y 2 ,..., y m ) (2)
Equation (9) to convert the QR to a numeric value, where
Then, we propose the fuzzy similarity matrix S XY between µ QR represents the membership degree, ν QR represents the
vector X and vector Y, in which x i y j represents the semantic non-membership degree and ρ QR is uncertainty degree.

– 128 –