九十七學年度下學期 類神經網路 研究計畫書

一、研究計畫中英文摘要:請就本計畫要點作一概述,並依本計畫性質自訂關鍵詞。(五百字以內)

 

 

  Customer Relationship Management (CRM) is a key element of modern marketing strategies.KDD CUP 2009 provides large-scale marketing databa from Fsesrance Telecom ORANGE company. Our purpose is able to predict customer's preferences. We are also able to provide then suggestion about surrounding Products and make them more profitably.We will use neural network analysis a large database.

 

So far of Neural network still have a new theory of architecture which has been raised constantly. Because the computer increase in computing capacity, it makes neural networks more powerful. Neural network has input layer, hidden layer and output layer.We can find hidden layer's relations from the expected input and output data. We use an enormous amount of data to calculus and a series of studing.

 

 

 It makes results of neural network will be more and more precise. The different algorithms will also affect  studys and results. We will able to create predictable platform from Neural Network that we prior had learned.The main characteristic of this platform is its ability to scale on very large datasets with hundreds of thousands of instances and thousands of variables. The rapid and robust detection of the variables that have most contributed to the output prediction can be a key factor in a marketing application.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

二、研究計畫內容:

Research background and motivation

     The main purpose create customer relationship management system and can Improve relations between customer and vendor. Vendor can understand the requirement about customers and can improve vendor sales rate. Vendor offers statistical data about customer and the database is too large . So we use neural network to learn. Neural network have always played an important role in research and application for many years. Their applicative areas include networking, investment, engineering, management and etc. The most commonly use algorithm in neural network which is the back-propagation(BP) learning.The more common structure is feed-forward back-propagation algorithm. The calculational process contains forward conduction and reverse transmission. The input signal  be treated via weight and then to  send to hidden layer.It can get output value through the transfer function. If the output value and expectations are not same , reverse transmission and correct network weights. It make error to reduce ,so it has function of learning and recall. It can do training and prediction.Because Neural model is to keep on being trained, relatively the accurate prediction value is high. It is to master customer of preferences and trends,and the market is one of the key.

 

Researchful method

(一)Neural network & algorithms

    Neural Network(NN) is similar to the structure of the human neural and it is base on study of brain and nervous system. It not only start with information technology but also has the basic characteristics of human brain : Learning, memory and induction. Neural network has the following types of neurons

 

processing element; PE

 summation functionactivity functiontransfer function

        

layer

 regular outputcompetitive output competitive learn

 

network

  learning course — supervised Learning NetworkUnsupervised Learning NetworkAssociation learning networkBidirectional associative learning networkoptimum learning network

 

recalling course —supervised recalling NetworkUnsupervised recalling NetworkAssociation recalling networkBidirectional associative recalling networkoptimum recalling network

 

Multi-input neuron

 

 
 


 

 

 

 

 

 

 

 

 

      output (a) = f ( WP + b )

The basic principles of Back-propagation network is to use the concept of Gradient Steepest Descent Method to minimize error function. That is, back-propagation network is used to let the output layer unit error, layer by layer to let the input layer back-propagation to different layer unit. At all layer obtain a referential error and it link to the weight to adjust the value of the corresponding. It can minimize the error between output value and target value. We suppose that  neural network's training are defined as follows typomorphic :

 

                                                                 

The input variables is x. The Target variables is t. The variables p is total number of training. If the type of neural network has been trained ,the actual output and target output are different. The back-propagation (BP) algorithm will be repeated through the data to train neural network. it adjust link's weights and bias weights until they reach the goal of set. Our goal is to minimize functional error as follows.

              

   The output layer variables  is . The variables  is a output layer of inferential value. The Back-propagation network is to use Gradient Steepest Descent Method to minimize error function. In order to reduce the amount of error, it must be of the opposite direction adjust link's weights of the network; Link-weighted adjust variable value of  relationship as follows:

                                                                                     

                                                                                        

The variables  is learn's rate. After we through  chain rate to calculate,link weights can be adjusted to change the amount of rewriting of the following results.

                                                           

                                                                                 

One of the  and  return value to layer, the complete expression is as follows.

                                                               

                                                                     

We can get the weighted express which contains the inertia link

The hidden layer is as follows

                                               

The output hidden layer is as follows

 

                                                                                                              

One of the inertia factor usually is value between 0.1 to 0.8 range.

Therefore, as long as I give neural network  a error tolerance , a studying the rate of , a inertia factor  and the largest study circle ,and then we can adjust the network by weight to achieve neural network for the customer data mining.

 

() researchful situation

we will be screening on variables because the variable is too many. SPSS process datas very well for the current and SPSS  have many statistical methods itself. We used SPSS to analysis each the relationship between variable and target. If the relation is much better , it will affect the output result. After treatment of the information,we use neural network to learn in NNTOOL of MATLAB. The calculated capability of MATLAB is a very strong software. It have many function itself ,so we writing the program more convenience. There are many algorithms in NNTOOL .Ddepend on the difference algorithm will effect the learn efficiently and prediction.We use the BP algorithm here and this is only a preliminary decision.

 

() Algorithm problem

   The input has 15000 variables ,the target is  only -1 and 1. The transfer function  use the hardlims. We discover that will produce problem to revise weigh.The output value and expectations will only appear -2,0 and 2. We get the information variables that is too large. Too many data amounts cause many softwares cant work , looking for new ways to solve this problem

 

Expected to achieve

   We will solve problem of database at present and the neural network has trained. The ultimate goal is to achieve accurate prediction.