HTML
-
The objective of fuzzy analysis is to categorize samples or variables with multiple attribute indicators in a certain way. There are two types of classification method: Q-type clustering that categorizes different sample objects with multiple attribute indicators and R-type clustering that classifies different attribute indicators of multiple sample objects. The clustering analysis adopted for the evaluation index of public meteorological services belongs to R-type clustering.
-
Multilayer clustering is required for samples containing different levels of evaluation indicators. In the proposed public meteorological service evaluation system, there are two levels of evaluation indicators suitable for clustering analysis. Multilayer fuzzy clustering analysis first classifies the features of the samples based on the degree of influence on the samples at different levels. Then, the weights of each feature are determined at each level and a fuzzy equivalence matrix is established for clustering analysis (Wang and Shu [47]). The expert evaluation method can be used to determine the weight of an evaluation indicator. Suppose there are M characteristics on level L. Then, we compare the importance of these M features, and either the expert evaluation method can be used to determine their weight vectors (ω11, ω12, ω13, …, ω1m) or another technical method can be implemented to determine their weights. Fuzzy equivalent matrices on the same level can be determined using corresponding fuzzy equivalent matrices and their respective weight coefficients:
$$ R_{i}=\sum\limits_{j=1}^{M} \omega_{i j} R_{j} $$ (1) where Rj is the fuzzy equivalent matrix corresponding to the j characteristics of the sample on level L. If the same, the final fuzzy equivalent matrix can be determined:
$$ R=\sum\limits_{i=1}^{n} R_{i} \omega_{i} $$ (2) where Ri is the fuzzy equivalent matrix corresponding to the i characteristic on the first level, and ωi is the product of the weights of each layer on the first level of i from the initial layer to the final one.
It can be seen that the principal idea of multilayer fuzzy clustering is to take into account the factors of the weight coefficients of each layer, and the total weight coefficients are the product of the weight coefficients of each layer. In this paper, the expert evaluation method is not used to determine the weight coefficient; instead, it is based directly on the importance of the data in the satisfaction survey. The weight coefficient is determined using the analytic hierarchy process.
-
The data used in this paper is from a public satisfaction survey of Wafangdian City in 2010. A total of 100 questionnaires were spread to collect data from employees of enterprises and institutions, civil servants and ordinary citizens. A total of 75 valid questionnaires were returned. Satisfaction survey involves three aspects: meteorological service information content, meteorological service information release and meteorological knowledge dissemination, each of which includes three or five questions. Wafangdian Satisfaction Survey results in 2010 are on the average level of the overall satisfaction of the whole country and thus are representative. Therefore, based on 75 groups of scoring results, multi-level fuzzy clustering analysis was carried out.
To facilitate rapid evaluation of the survey data, for each item in the questionnaire, the respondents were requested to select one of five possible responses: "Satisfactory", "Reasonably satisfactory", "Adequate", "Unsatisfactory"and"Not sure". The numeric scores corresponding to these five options were 100, 80, 60, 0 and average scores. Therefore, the first step in the treatment of the original data was to transfer all 75 sets of data into specific scores, which are the basis for a data matrix.
-
R-type clustering differs from Q-type clustering because the classification is performed according to different attribute indicators of the sample objects rather than different samples with multiple attributes. Thus, the row-column construction of the data matrix in R-type clustering is opposite to that of Q-type clustering. To perform R-type clustering for the evaluation index of meteorological service satisfaction, the 13 elements of the second-level evaluation indicators shown in Fig. 1 were used as matrix rows and the 75 sets of evaluation data were used as matrix columns.
To undertake multilayer fuzzy clustering, the weight coefficients of each layer should be determined first. In this paper, analytic hierarchy process (AHP) is used to determine the weight of meteorological service evaluation index adopted by Yan et al. [48]. Through the steps of constructing judgment matrix, matrix consistency test, and determining the maximum eigenvalue, i.e. weight vector, the weight coefficients of the first and second indexes are finally obtained as shown in Tables 1 and 2, respectively. Indicators 1-13 in the first row of Table 2 represent the 13 secondary evaluation indicators illustrated from left to right in Fig. 1.
First-grade evaluation indicators Meteorological service information content Meteorological services information release Meteorological knowledge dissemination Weight coefficients 0.1634 0.2970 0.5376 Table 1. Weight coefficients for first-level evaluation indicators.
Indicators 1 2 3 4 5 6 7 8 9 10 11 12 13 Weight 0.1267 0.2972 0.4380 0.0798 0.0583 0.7703 0.0679 0.1618 0.2213 0.0628 0.0698 0.2645 0.3817 Table 2. Weight coefficients for the second-level evaluation indicators.
The total weight coefficient of the target layer was obtained by multiplying the weight coefficients of the corresponding first-level and second-level evaluation indicators, from which the final score of the 75 groups was obtained. Thus, the construction of the raw data matrix was completed; however, because of the sheer volume of data, there is no specific data matrix.
-
Various dimensions can be used to describe the characteristics of things. Therefore, to eliminate interference in the calculation process, the matrix must be standardized. There are two principal methods for standardization.
The first method is the translation of the standard deviation, which can be implemented by using the following formula:
$$ {{\rm{x}}_{ij}}^\prime = \frac{{{{\rm{x}}_{ij}} - {{\rm{x}}_i}}}{{{{\rm{s}}_j}}}(i = 1, \ldots , n, j = 1, \ldots , {\rm{m}}) $$ (3) where
$$ \bar{x}_{j}=\frac{1}{n} \sum\limits_{i=1}^{n} x_{i j} $$ (4) and
$$ s_{j}=\sqrt{\frac{1}{n} \sum\limits_{i=1}^{n}\left(x_{i j}-\bar{x}_{j}\right)^{2}} $$ (5) The second method is the translation of displacement-extreme difference, for which the formula of transformation is as follows:
$$ {{\rm{x}}_{ij}}^\prime = \frac{{{{\rm{x}}_{ij}} - \min \left\{ {{{\rm{x}}_{ij}}\mid 1 \le i \le n} \right\}}}{{\max \left\{ {{{\rm{x}}_{ij}}\mid 1 \le i \le n} \right\} - \min \left\{ {{{\rm{x}}_{ij}}\mid 1 \le i \le n} \right\}}}(j = 1, \cdots , m) $$ (6) The method of translation of displacement-extreme difference was used in this study for standardization during data processing.
-
For the standardized processing matrix above, the degree of similarity among the classification objects was calculated to establish a model similarity matrix (Matrix R= (rij)n × n), which is a process also known as calibration. After data standardization, the matrix was transformed into an n-order array after calibration, where the elements in the matrix fall within the specified range. Various methods can be used to calculate and demarcate the model and the similarity matrix, and many methods are available for the construction of the model and the similarity matrix. In this study, we used the Euclidean distance method:
$$ r_{i j}=1-c \sqrt{\sum _{k=1}^{m}\left(x_{i k}-x_{j k}\right)^{2}}, i, j=1 \ldots \mathrm{n} $$ (7) where c is a constant to ensure 0≤ rij ≤1.
-
The final model and aggregation analysis can be produced using various methods, i. e., (1) clustering methods based on the modular equivalent matrix, including the transfer closure method and the Boole matrix method; (2) direct clustering methods based on model similarity, which include the maximum tree method and the network method; (3) the modular clustering method based on the model c-mean value. In this study, the method used most widely for solving the model and the equivalent matrix, i.e., Matrix tr(R), was used. Given R2=R·R, R4= R2·R2, and R8= R4·R4, i.e., R→ R2→R4→R8→…R2k, then after such a limited operation, when R2k= Rk is available, Matrixtr(R)= Rk, which is the equivalent modulus of Matrix R, can be considered transitive. Thus, different truncated fuzzy matrices can be obtained by choosing the appropriate confidence level [0, 1], i.e., the results of fuzzy clustering analysis at different levels of λ.
-
All the above steps involved large amounts of matrix data calculation, which was performed conveniently using the Matlab program. The main ideas and methods were as follows. First, the main program file jl.m was established. This not only specified CS = 2, i. e., the use of translation of displacement-extreme difference for data standardization processing, and used the original data A established in"2.4.1, "as input, but also cnamed three. m files: F_JISjBzh. m, coesimeu. m, and F_JlDtjl. m. Second, the. m file (F_JISjBzh. m) of standardized data was established. Third, the.m file for the constructed fuzzy similarity matrix (coesimeu. m) was established. Fourth, the. m file for the fuzzy clustering graph (F_JlDtjl. m) was established. The specific codes in each.m file are omitted here.