Taguchi's T-method is a method used to predict an output value from multivariate data. The T-method has the advantage that calculation is simpler than multiple regression analysis. In time series analysis, it is important to reduce the amount of calculation. In this paper, we propose a method of applying the T-method to a univariate time series analysis. Moreover, we propose simulating this under various conditions. For prediction accuracy, we compare the T-method, the fit of the AR model, and improvement T-methods proposed in the previous research. Improvement T-methods are Ta-method and Tb-method. This study shows that prediction using T-methods is more accurate than prediction using an AR model under certain conditions. When time series data to be analyzed is close to a nonstationary model, the accuracy of prediction using the AR model worsens. This is because the AR model assumes stationarity but, Taguchi's T-method does not. As a result, we can show that using a T-method for univariate time series analysis is robust against the target model and autoregressive coefficient value. That is, in this method, there is little decrease in prediction accuracy owing to the difference in the nature of time series data.

View full abstract

Analysis of customer purchase behavior is an active research area. Several recent studies in this area have focused on detecting groups based on similar features from the available business data. One of the efficient approaches used for dividing all data into several groups with similar features is called clustering. Application of clustering to customer or item data can generate results that can help a company develop efficient marketing strategies. However, it is a non-trivial task to apply a general clustering method to a dataset with high dimensionality and sparsity owing to the difficulty of defining an appropriate similarity metric between the data samples. In this situation, one should reduce the dimensionality of data in advance. The autoencoder is a well-known model based on neural network, which can convert high-dimensional nonlinear data to low-dimensional expressions. In this research, we propose a method for clustering high-dimensional sparse data efficiently with deep learning method. The first step of our proposed method is to adopt SDAE for dimensionality reduction. The second step is to conduct a clustering analysis using the hidden layer expressions of SADE. The effectiveness and the usability of the proposed method are clarified through its application on real purchase history data.

View full abstract

Along with the recent developments in information technology, purchasing on E-commerce (EC) sites has become popular and Web marketing measures are growing in significance. However, conversion rates on EC sites are usually not high, so there is room for improvement in Web marketing. In general, a user browses each page on an EC site before he/she purchases an item, and in many cases, the user leaves the EC site without purchasing. Therefore, constructing a binary classification model with browsing paths is considered here to obtain some findings for an effective Web marketing strategy. This model learns the customers' browsing history data with their purchase and non-purchase labels that are accumulated on an EC site. It is generally considered that there are different page transition tendencies between purchasing and non-purchasing users on an EC site. This study focuses on the difference in page transition patterns between purchasing and non-purchasing users. To model customers' page transition behaviors, N-gram and mutual information are introduced to select features from browsing history data. In addition, a binary classification experiment using real browsing history data was carried out and the effectiveness of the proposed method is demonstrated.

View full abstract

Statistics are indispensable for understanding how industries are structured and for developing business strategies. Here, we aim to improve the estimation accuracy in the statistics for the estimated scale of the foodservice market. First, three selected statistical surveys of the foodservice industry were evaluated to determine if they can be used for a new estimation method. Second, a new method was proposed for estimating the scale of the foodservice market. Finally, we assessed the accuracy of the missing-value supplementation in four types of estimation methods. Our results showed that the sales from JF yearbook, a survey of the foodservice industry by Nikkei's Marketing Journal, and the Basic Survey on Small and Medium Enterprises can all be used for estimating the scale of the foodservice market. Moreover, by combining the data from the three surveys and conducting a limited additional survey of the enterprises not included in these surveys, a more reliable estimate for the scale of the foodservice market could be obtained. Finally, it was found that the random forest algorithm was the best unified supplementation method; the highest accuracy can be obtained by applying the most suitable method for each industry and combining the results.

View full abstract

In tournament systems, round robin tournament and elimination tournament are different methods for facilitating and organizing the games while featuring competitions which involve only two participants per match, and they are used prominently for ranking participants. However, each tournament system has disadvantages, such as the time required to complete the tournament and difficulties in ranking all participants. To compensate for these disadvantages, our previous study proposed a new tournament system that uses a paired comparison methodology based on the concepts of cyclic designs and Swiss system tournaments. We focused on the relationship between the rank estimated by each method and the true strengths of players while comparing the accuracy of each tournament system ranking and the number of games required to complete each tournament system.
In this study, we add a new evaluation index that accounts for the relationship between estimated rankings and the game results, such as the number of wins and acquired sets. We indicate that the rank estimated using our proposed method reflects not only the game content but also the game results. Moreover, the actual Elo rating value is used to establish the strength of players. The usefulness of the proposed method is verified by reproducing a more realistic situation.

View full abstract