Professional Documents
Culture Documents
Method On Determining Feature Weight in Case-Based Reasoning System
Method On Determining Feature Weight in Case-Based Reasoning System
Linear transfer function, the expression of function as The feature weighting value is 1 after standardization.
follows: According to the formula (6) and (7), calculations of the
§ 0 i characteristics of the weight.
¨ | p − p | ·¸
¨
¨
¦ L 0 ¸
¸ B. Theorem
p
si = © ¹
Theorem 1 Decision-making to the characteristics of
n ˄˅
the target data set composed of samples of the case library
S-shaped hyperbolic tangent transfer function, the
as a BP neural network learning samples for training, For
expression of function as follows:
Si and Sj , respectively, for the deletion of the first items i
si and j after the characteristics of the sensitive coefficient.
wi = n IF Si>Sj, the sensitivity of the feature i larger than the
¦
j =1
s j sensitivity of feature j, that is Wi>Wj.
˄˅ Theorem 2 Decision-making to the characteristics of
the target data set composed of samples of the case library
S-shaped logarithmic function, the expression of
as a BP neural network learning samples for training, Pi is
function as follows:
assumed as predictive value of network output after
0 i
lni =¦
p −p| | deleting feature i,P0 is assumed as predictive value of
L 0 network output without deleting any nodes, feature
p ˄˅ weight vector denoted as {l1,l2,…ln}.
0 i
III. SENSITIVITY ANALYSIS | p −p|
A. Definition
lni = ¦L 0
Authorized licensed use limited to: McGill University. Downloaded on February 16,2023 at 01:40:43 UTC from IEEE Xplore. Restrictions apply.
IV. DETERMINING METHOD ON FEATURE WEIGHT Step6: Calculate network sensitivity coefficient after
A. Determining Method on Feature Weight Based deleting an input node, then in accordance with formula
on BP Neural Network and Sensitivity Analysis (7), calculate the feature weight.
Data sets are usually given in the form of tables,and Step7: Repeat to perform steps 6, according to
we call a case system, formulated as T=<U,A,V,,f>, different features of the case, we delete different input
where U is a set of finite and nonempty objects called the nodes from the network, that is, delete the various
universe,A is the feature set,V is the domain of the feature features of the case one by one, calculate their sensitivity
value,and f is the information function:f =U×AĺV.If the coefficients, and so all features weight are determined.
attribute set A is divided into condition feature set C and B. Cases Analysis
decision feature set D.The structure of cases accord with In this section, an example is used to illustrate the
people's habits and is good for the storage and neural analysis procedure proposed. In Anhui Agricultural
network training.For studying the dynamic relationships University red-seed watermelon horticulture 150 cases
of these features, in this paper,we construct model of BP are randomly selected as training data, 80 cases are
Neural Network,where feature set C are regarded as input chosen as testing data[8].It can be regarded as a set of
nodes of BPNN and feature set D are regarded as output attribute values of the feature of (feature value pairs)
nodes of BPNN.We analyse the degree of sensitivity of constitute a collection. In the case, there are many
the output neurons to the input neurons,when input characteristics can be described about red-seed
neurons are deleted one by one.As nodes reduce and watermelon. According to expert experience some
weights among neurons change,on the corresponding the features were concerned, such as fruit weight, seed size,
topology of the network may dynamic change. seed number, grain weight and seed weight and other
Then determining algorithm on feature weight based characteristics, as a division of the feature on different
on BP neural network and sensitivity analysis is types of red-seed watermelon kernel.
constructed as follows: In accordance with the above algorithm, by the
Step1˖Firstly, Select representative data as a sample training neural network model forecasting result is as
case data in case-based system; shown in Fig.1, which meets the requirements on
Step2˖ In order to reduce the effect resulting from the prediction accuracy.
different condition feature set, we normalize the
numerical attribute x into the interval [0, 1] with
xk −xmin
xk' =
xmax −xmin ˄˅
Step3˖Determine the network topology, input neurons
and output neurons are determined in different cases.
According to the formula (1) or (2) to determine the
Figure 1. The training neural network model forecasting results
number of hidden layer neurons;
Step4˖Select the learning parameters of BP Neural After the dynamic adjustment of neural network
Network; determine the network's transfer function and structure on testing data, by the training network different
training function, training steps and target errors. models of BP Neural Network prediction error curve are
Step5˖Training of the neural network to determine the shown in Fig2,where y1, y2, y3, y4, y5 representatives of
accuracy of the network to meet for the purpose of deletion fruit weight, seed size, seed number, grain weight
training the network. If the network does not meet the and seed weight. Between the real data and the
accuracy, then we go to Step 3 and repeat to perform steps forecasting data, which show the BPNN-SA model is
3 and step 4.Optimization of neural network parameters valid in precision of forecast.
and structure, adjust the training function,re-learning on
the network until meet the given accuracy. Otherwise, go
to Step 6.
Authorized licensed use limited to: McGill University. Downloaded on February 16,2023 at 01:40:43 UTC from IEEE Xplore. Restrictions apply.
The calculation results of weight are as shown in table
1. We can conclude that the feature of the seed size is the
smallest weight in all features. So the contribution of
weight on seed size is the smallest. Other features weights
followed by descending: fruit weight, seed number, grain
weight, seed weight. The results are in line with experts
and practitioners. That is to say, the feature with high
weight play important role in decision classification and
prediction.
Feature Weights
Algorithm
fruit weight seed size seed number grain weight seed weight
V. CONCLUSION REFERENCES
[1] Jiang Yan,Chen Rongqiu.Parameter Sensitivity Analysis Based on
Therefore, we can conclude that the algorithm of the Least Variance of Weights. Journal of Wuhan Automotive
BPNN-SA not only reflect the dynamic relationship Polytechnic University. vol.27,NO 1, Feb. 2005, pp.165-167.
[2] Zhu Yin,Meng Zhiyong,Kan Shuyu.Determination of Weight V
between features, but also calculate simply and overcome alue by AHP. Journal of Northern Jiaotong University. vol.23,NO
the one-sidedness. So it is accurate and effective. The 5, Oct. 1999,pp.119-122.
[3] Walczak B, Massart D L. Rough sets theory. Chemometrics and I
algorithms provide a basis for retrieval and re-use in ntelligent Laboratory Systems, vol.47,NO 1, Apr,1999, pp.1-16.
case-based reasoning system. Our future work includes the [4] SUN Huijun,WANG Xinhua.Determin ation of the Weight of
Evaluation Indexes with Artif icial Neural Network Method.
extension of determining method on feature weight. Journal of Shandong University of Science and Technology.
vol.20,NO 3, Sep. 2001,pp.84-86.
VI. ACKNOWLEDGEMENTS [5] Wang Wenjian.Improve of BP Neural Networks Model. Computer
Engineering and Design. vol.26,NO 11, Nov. 2000, pp.8-10.
The authors acknowledge support from the National [6] Zeng, X., & Yeung, D. S.Sensitivity analysis of multilayer
perceptron to input and weight perturbations. IEEE Transactions
Science Foundation of China (Grant No.30800663 and on Neural Networks, vol.12,NO.6, Dec. 2001, pp. 1358-1366.
[7] Zeng X,Yeung DS.Hidden neuron pruning of multilayer
Grant No.30971691). perceptrons using a quantified sensitivity measure.Neuro
computing, Vol. 69, Mar.2006, pp. 825-837.
[8] Yang Baohua, Li Shaowen, Fan Jianfeng. Distinction method for
red-seed watermelon quantitative trait based on rough set
.Agriculture Network Information, vol.6,Jun.2006, pp.38-39,58.
Authorized licensed use limited to: McGill University. Downloaded on February 16,2023 at 01:40:43 UTC from IEEE Xplore. Restrictions apply.