電子商務期末報告 報告人：杜孟儒 電話 : 03-5915800/0955948239 學號： 9534805( 博一 ) 指導老師：蔡銘箴教授 日期： 96 年 7 月 3 日

1. Predicting and explaining patronage behavior toward web and traditional stores using neural networks: a comparative analysis with logistic regression(Decision Support Systems 02/2004) 電子商務期末報告 報告人：杜孟儒 電話: 03-5915800/0955948239 學號：9534805(博一) 指導老師：蔡銘箴教授 日期：96年7月3日

2. 報告大綱 • 研究動機(Motivation) • 文獻探討(Background and related works) • 研究方法學(Methodology) • 實驗結果與比較(Empirical Results) • 檢討與建議(Comment and Suggestion) • Q & A

3. 研究動機 • Developing neural network models to predict and explain consumer choice between web and traditional stores • Proving neural network models can significantly outperform logistic regression models in predicting power regarding consumer’s patronage behavior • Identifying factors that have significant impact on customers’ channel attitude

4. 文獻探討(1) • Very crucial in this E-Commerce age to understand consumers’ attitude toward web stores. • Most studies seeking to address the questions regarding customers’ channel choice are based on empirical surveys and statistical analyses. • Most surveyed empirical studies are forms of “preference regressions and they all share the same a priori assumption that the process of consumers’ channel evaluation is a linear compensatory. • Compensatory: Assume that any shortfall in one channel attribute (e.g., immediate possession of a product) can be compensated by enhancements of other channel attributes (e.g., price).

5. 文獻探討(2) • Consumers’ behavior might not always be compensatory. For instance, consumer concern may just be immediate possession of a product and this concern may not be compensated by the enhancement of other channel attributes, such as price. • Compensatory statistical models may not be able to capture non-compensatory decision rules. • Do non-compensatory choice models using neural networks perform better than logic choice models in predicting consumers’ channel choice between web and traditional stores ? • If so, based on the non-compensatory choice models, what are the main predictors of customers’ online buying behavior ?

6. 研究方法學 • Overview of neural networks model • Consumer survey methods and procedures • The logit model of consumer channel choice • The neural network model of consumer channel choice • Neural network training • Sensitivity analyses

7. - mk x0 w0 x1 w1 f å output y xn wn Input vector x weight vector w weighted sum Activation function Overview of neural networks model: The structure of a neuron • The n-dimensional input vector x is mapped into variable y by means of the scalar product and a nonlinear function mapping and represents the bias of the unit. The bias acts as a threshold in that it serves to vary the activity of the unit

8. Overview of neural networks model: A Multi-Layer Feed-Forward Neural Network Output vector Output layer Hidden layer wij Input layer Input vector: X

9. Survey methods and procedures:Overall survey procedures

10. Survey methods and procedures:Questionnaire Design

11. Survey methods and procedures:Selection of channel attributes Used Economic Factors to access customer’s attitude toward web stores shopping. Opportunity Cost Search Cost Uncertainty Examination Cost Comparison Cost Asset specificity Payment Cost Used transaction cost analysis (TCA) for explaining of online consumer behavior. A list 18 attributes that may affect a customer’s decision to purchase from web were chosen based on TCA. Delivery Cost Post-Service Cost

12. Survey methods and procedures:Product Categories Six products used in the survey and their characteristics

13. Survey methods and procedures:Survey Outcome Attribute Performance ………… For product j, web stores were perceived to have a higher level of attribute i than traditional stores For product j, web stores were perceived to deliver a lower level of attribute i than traditional stores

14. The logit model The relative utility of a consumer purchasing product j from a web store is defined as (4 is the midpoint of the scale)

15. The neural network model : A Three-Layer Back-Propagation (BP) Neural Network • Six separate BP networks, one for each individual product • In each BP networks, the input layer consisted of 18 nodes, each corresponding to 1 of the channel attributes • The output layer had a single node with two values representing the consumer’s choice (either Web (1) or traditional (0) stores) Consumer choice (1 / 0) channel attributes n = 1 m = 18

16. The neural network model : A Single Neuron j in the Hidden Layer • Three nodes in the hidden for DVD and books • Four nodes in the hidden layer for the rest of products a(n) 1 1/2 0 n

17. Neural network training:The procedure of network pruning

18. Sensitivity analyses(1/2) • A sensitivity analysis is used to measure the response of the network to the perturbation of neural network parameters such as inputs and weights. • It measures the effect of altering the value of an input variable (e.g., channel attributes) on the output value (e.g., patronage behavior). • The channel attributes that have little or no impact on the prediction of patronage behavior will produce low sensitivity values; such attributes are considered insignificant and should be removed form a neural network model. • A reduction in the number of input variables directly decreases the total number of feed forward and backward propagation calculation. • Such optimization offers advantages in terms of simpler networks, faster training and better generalization ability to avoid overfitting due to the oversized network • Pruning a neural network by removing insignificant input nodes may increase the predictive accuracy

19. Sensitivity analyses (2/2) where yj=output of hidden node j ( j=1, 2, . . ., l); Wji=weight of the connection between the hidden node j and the input node i; Wkj=weight of the connection between the output node k and the hidden node j. A change in due to a perturbation implies a change in F(sum). Given the same perturbation to Xi, higher sensitivity is achieved when the change in F(sum) is larger

20. 實驗結果與比較: Comparison of predictive performance Predictive accuracy of neural networks and logistic regression • Neural networks, their optimization process resembles the minimizing of the error term in that of standard regression • Neural networks consider linear, non-linear and pattern recognition relationships in the input data and conduct the optimization process automatically • Neural networks use a unique algorithm in such a way that the technology does not have a problem with multicollinearity, which can cause major errors in standard regression analysis

21. 實驗結果與比較: Finding the drivers of consumers’ channel attitude (1) Sensitivity of output with respect to channel attributes for books

22. 實驗結果與比較: Finding the drivers of consumers’ channel attitude (2) Attributes affecting consumers’ channel choice based on sensitivity analyses

23. 檢討與建議

24. The proposed new decision support model based on fuzzy neural network (FNN) and Genetic Algorithm Attributes determination for consumer patronage behavior toward web and traditional stores Questionnaire design and survey for qualitative factors based on Fuzzy Delphi method Fuzzy IF-THEN rule base development Generating the initial weights via Genetic Algorithm Fine-tuning the FNN via error back-propagation (EBP) type learning algorithm Consumer patronage prediction model

25. The structure of fuzzy neural network (FNN)

26. Q & A

27. Next • A recommender system using GA K-means clustering in an online shopping market(Expert Systems with Applications, 2007) • A decision support system for order selection in electronic commerce based on fuzzy neural network supported by real-coded genetic algorithm (Expert Systems with Applications, 2007)