Many companies need to know how to appeal to consumers in favor of their products and services by broadcasting commercial messages (CMs) through TV. CMs are expected to create and nurture the best image of the company and the products among consumers, and to promote the purchasing activities of consumers through consumer viewing. However, these effects are difficult to produce without strategies. CMs should be designed with the appropriate concept and content. Therefore, there is a need to recognize what stimulates the psychological change in consumers leading to purchase activity: that is, the physical change of recognizing the item and buying the item. Subsequently, and the element(s) of the physical change should be analyzed to shed light on the appropriate target consumers for designing CMs.
In this study, we focus on a popular yogurt brand in Japan called “Meiji's Purobio yogurt R-1”, analyzing the psychological change in consumers attributed to CMs on TV based on the attributes of 3,000 consumers' data and their TV browsing history. For grasping the psychological state of consumers, we apply the attention, interest, desire, action, satisfaction model and define the consumers' five psychological states based on questionnaire data. The factors of CMs with effects on consumers in each psychological state are clarified using two clustering approaches.
Now that it is possible to handle the enormous amounts of data available on the Internet and generated by corporate systems, automatic classification technology has assumed greater importance. Various learning methods have been proposed for solving two-class classification problem. Furthermore, an ensemble learning method that uses a combination of a plurality of classifiers can deliver high accuracy. However, the imbalance of class labels, that is, the sample size between classes to be classified is greatly different, often occurs. It is reported that the ensemble learning method has poor accuracy in regard to such imbalanced data.
Therefore, in this research, we propose a novel analysis procedure based on ensemble learning for such imbalanced data. Our proposed procedure involves dividing data into several legions by clustering and uses the over-sampling technique to make data ease imbalanced state and learns classification rule based on the random forest method proposed by Breiman (2001). Through simulations, we show the effectiveness of the proposed procedure.
Control charts are a representative method of statistical process control. Control charts make it possible to effectively manage the manufacturing process but it assumes the availability of large historical data sets. In the high-mix, low-volume production environment that has become a mainstream in recent years, sufficient samples for estimating process parameters cannot be obtained often. In such a situation, the control chart does not function properly. In addition, in manufacturing processes such as cutting, process averages can change due to deterioration even if the process operates in the in-control state, leading to type I error increases. Therefore, herein, we use hierarchical Bayesian modeling to propose control charts that function appropriately for trendy data sets in high-mix, low-volume production. We then show the usefulness of such an approach by comparing with a conventional method. Since hierarchical Bayesian modeling makes it possible to assume the same distribution for parameters of various types, it is possible to use all the information to estimate parameters comprehensively. This capability makes the new approach effective for high-mix, low-volume production.
Plackett Burman designs, that is one of the fractional factorial designs, have a property that the columns to be assigned the main effects are mutually orthogonal. Plackett Burman designs have been used in screening experiments for identifying active main effects out of candidate factors frequently. The screened results are not always correct due to the experimental errors and the alias relations between main effects and two-factor interactions. While several previous researches has been discussed the alias relations in Plackett Burman designs, not many researches has been focusing on the alias relations between main effect and interaction in N=20 as a structured manner. In fact, although Hamada and Wu (1992) enumerated large combinations of correlations, this study has organized the structure manner into a clear and unique table as table4. We demonstrated the alias relations in a systematic manner as a form of 0.6-triplet, indicating that the correlation coefficient between estimates of main effect and two-factor interaction is equal to 0.6. Furthermore, a new rule to describe appearance of 0.6 triplet is demonstrated as an application of BIBD. We also propose a new assignment method in N=20 Plackett Burman designs that avoids 0.6-triplets. The assignment method is confirmed that it has better property comparing to the conventional approach.