In the original Bayesian network model, the hierarchical structure of the variables is not assumed. When modeling the relation between the sales of products in a retail industry, it is better to consider a hierarchical structure of items (e.g., first, second, and third classifications). To apply a Bayesian network to such data, we have to focus on one hierarchy only in order to acquire a Bayesian network model. However, focusing only on the first or second classification provides a high-level view, but makes it difficult to understand customers' purchasing behavior in detail. On the other hand, focusing on the third classification results in a considerable number of nodes and a complicated network structure. Thus, capturing the overall relationship between products is not straightforward. Therefore, we propose a hierarchical Bayesian network model and a new learning method based on max-min hill-climbing learning algorithm. In the proposed method, we focus on the hierarchical structure of products, which enables us to construct a lower layer that considers the causal relationships in the upper layer. Furthermore, we investigate a case study using actual hierarchical data on consumer purchases, and demonstrate the proposed model using a simulation analysis.
In recent years, many companies conducting recruitment activities and many students looking for a job use internet portal sites for job-hunting in Japan. Companies can post their basic information on individual company pages and recruit applications from students. On the other hand, student users can gauge corporate attractiveness by browsing individual company pages on a job-hunting site and can make entries to companies of interest. Therefore, a large amount of their behavior history data is accumulated on the site. There are several studies on prediction of users' entries to companies and analysis of preference using user attribute information and entry history data. However, in the conventional researches, browsing activities on individual company pages existing in the background of the user's entry were not considered, so the relation between browsing and making entries has not been studied. This research proposes a latent class model for analyzing the relation between browsing company pages and making entries to companies. The proposed model enables clarification of target users and consideration of effective promotion activities. Through a demonstrative analysis using actual data on a major job-hunting website in Japan, we show the effectiveness of the proposed model.
The Japanese nation has experienced a variety of disasters. It is therefore important for hospitals to function as social infrastructures through the Business Continuity Plan, and to construct a Business Continuity Management System (BCMS). To construct a BCMS, Risk Assessment (RA) needs to be conducted. This is necessary to identify, analyze, and evaluate risks that cause business interruptions. Additionally, from the RA results, countermeasures can be created to improve business continuity during disasters. However, the methods of RA and countermeasures planning to be implemented of hospitals during a disaster have not yet been clarified yet. Therefore, in this study, we propose a method for hospitals to conduct RA and countermeasures planning.
First, we ensured medical staff at a hospital implemented RA and identified risks. Second, we analyzed the problems for each RA procedure and focused on two major problems: inability to identify risks and inability to take countermeasures. To solve these problems, this study developed two tools. First, we gathered data on the issues hospitals face that are caused by disasters and created a risk list. This enabled medical staff to identify risks more easily. In addition, we created a countermeasures viewpoint list. This assisted medical staff in taking countermeasures.
Periodic inspections can accurately detect hidden failures. They are called perfect but are usually costly. The periods of perfect inspections (PIs) cannot be easily changed due to regulations and management problems. With improvements in monitoring technology, many devices for remote inspections are available at reasonable cost when compared with those for PIs. Remote inspections are called imperfect in that hidden failures can be detected with a probability of less than 100%. The periods of imperfect inspections (IPIs) can be easily adjusted. We propose an integrated inspection policy that combines PIs and IPIs and is flexible in that the frequencies of IPIs need not be common for all PI periods. The inspection frequencies of IPIs are determined to minimize the total expected cost, which includes inspections and penalty cost due to undetected failures. We investigated the property of the total expected cost function for optimization of IPI frequencies and used this obtained property to implement an optimization algorithm. We provide numerical examples to compare the expected costs of the proposed policy with flexible IPI frequencies and a policy with common IPI frequencies. These examples indicate that the proposed policy can reduce the total expected cost for systems with increasing failure rate and enhance the effectiveness of the inspection plan.
In Japan, Specific medical examination is conducted for national health insurance subscribers of 40 to 74 years old to prevent lifestyle-related diseases. However, problems remain in effective utilization of these examination results. In this research, we classify health condition types using SOM, one of clustering method, from the Specific health examination results of Izumi City, Osaka Prefecture. We also examined how these types change. As a result of SOM, health conditions were divided into 20 nodes, and it was found that these nodes are characterized by blood pressure, blood glucose level and so on respectively. We divide these nodes into three patterns: healthy, boundary and metabolic. The metabolic pattern was found to have higher medical expenses and medical treatment rates than healthy pattern. In addition, it was found that each pattern transits via the boundary pattern during the 6 years, i.e. data collection term. Furthermore, the medical cost has been on an upward trend at any node, however we can confirmed the node which has transition to a node with a low medical expenses, and in these node, increase in medical expenses are suppressed. Our future work is how to reduce medical expenses by human intervention such as health classroom.
In retail businesses, it is important to grow customers into good customers in order to increase and maintain sales. Therefore, many researches have clarified the differences of the purchasing tendency between good and other types of customers; however, in fact, it is hard to say that many retailers can take useful measures for acquiring good customers effectively by taking into account the purchase tendency. On the other hand, a latent class model has been studied as an analysis method to grasp customers' preferences and purchasing trends from purchasing history data. In this study, we assume latent classes behind customers' purchasing data, and apply the Latent Dirichlet Allocation model to represent the differences in customers' preferences. This model enables grasping the tendency of purchasing items in each latent class considering diversity and heterogeneity of the features of customers and items. Moreover, we propose a method to extract important items for each customer in terms of turning him/her into a good customer based on the k-nearest neighbor algorithm by using the analysis result of purchasing trends in each latent class. Furthermore, we apply the proposed model to actual purchasing history data and show the possibility of its application in practice.