A hybrid model for optimizing customer preference predictions in product design
Quality improvement has been an industry objective since the beginning. Although many definitions of QFD exist today, the major trend is to look at it as an overall measure of goodness. The work proposed here tries to improve quality from the customer point of view. ^ Quality Function Deployment is one of the most widely used tools in the industry to capture customers' requirements and incorporate them into the product design in order to guarantee a desirable product for the potential users. This work studies QFD and the introduction of conjoint analysis within its framework, proposes improvements to the model, and implements a new solution that includes all these improvements. ^ Three hypotheses are presented that aim at verifying the improvements made to the tool. The three hypotheses correspond to the three main improvements proposed in this work: ^ The first hypothesis postulates that changing the expert information elicitation method from a “relationship strength” method to a “frequency based” method will increase the model accuracy. In order to test this hypothesis two QFD models were created: The Traditional and the Frequency based QFD models. These models were named TQFD and RQFD respectively. The tests show that RQFD has a 5% higher accuracy than the TQFD and that these results are statistically significant at an alpha level of 0.05. The investigators also believe that this difference is of very high practical significance, especially for the specific domain the experiments were performed in: The digital image quality perception domain. This improvement in the accuracy can influence the complicated image and video compression and coding techniques to become more efficient, create smaller files and keep the quality perception untouched or even improve upon it. ^ Second hypothesis proposes that incorporating correlation data pertaining to the engineering characteristics as well as the customer requirements will improve the model's accuracy. This hypothesis was tested using an additional model that took advantage of the correlations data. The model (named CQFD) was compared against the TQFD and showed a 4% improvement in accuracy which was again statistically significant at a 0.05 alpha level. This improvement is again of practical significance due to the same reasons mentioned in the previous paragraph. ^ The third QFD improvement is hypothesized to be realized if a conjoint analysis step is performed based on the QFD results. The connection between the two models is developed in this work and a hybrid QFD model (HQFD) is compared against the RQFD (best QFD model to this point) as well as the CRQFD model which incorporates both frequency and correlation changes shown in the previous two hypotheses. The results show that HQFD is about 2% more accurate than CRQFD which is a validation of the third hypothesis at 0.05 alpha level. The practical significance of this improvement may not be as evident as the statistical significance at first glance. However, the author would like to point out that this improvement will only be achieved in addition to the improvements made through the introduction of frequency based relationships and the inclusion of correlation data. The three improvements will not have an additive effect, but their combined results are more accurate than any single one (with the exception of the last improvement as it could not be tested in this study). This observation shows that the practical significance of the full HQFD model is even more than that of the two previous models. ^ The investigators also show (through a final experiment) that performing a QFD analysis step prior to a conjoint analysis has a positive effect on the conjoint analysis results. This test is limited in the current study and only shows a statistical significance in conjoint analysis accuracy improvement when a shortened conjoint base is created using QFD rather than a random choice. ^ This work opens the door to very promising future studies in QFD analysis and conjoint analysis as well as implementation of both tools in the image quality perception studies and techniques.^
Mark R. Lehto, Purdue University.
Business Administration, Marketing|Engineering, Electronics and Electrical|Engineering, Industrial