Fumiaki Saitoh

On generalization error of self-organizing map

Self-organizing map is usually used for estimation of a low dimensional manifold in a high dimensional space. The main purpose of applying it is to extract the hidden structure from samples, hence it has not been clarified how accurate the estimation of the low dimensional manifold is. In this paper, in order to study the accuracy of the statistial estimation using the self-organizing map, we define the generalization error, and show its behavior experimentally. Based on experiments, it is shown that the learning curve of the self-organizing map is determined by the order that are smaller than dimensions of parameter. We consider that the topology of self-organizing map contributed to abatement of the order.

Extraction of Minority Opinion Based on Peculiarity in a Semantic Space Constructed of Free Writing

Recently, the “voice of the customer (VOC)” such as exhibited by a Web user review has become easily collectable as text data. Based on large quantity of collected review data, we take the stance that minority review sentences buried in a majority review have high value. It is conceivable that hints of solution and discovery of new subjects are hidden in such minority opinions. The purpose of this research is to extract minority opinion from a huge quantity of text data taken from free writing in user reviews of products and services. In this study, we propose a method for extracting minority opinions that become outliers in a low-dimensional semantic space. Here, a low-dimensional semantic space of Web user reviews is constructed by latent semantic indexing (LSI). We were able to extract minority opinions using the Peculiarity Factor (PF) for outlier detection. We confirmed the validity of our proposal through an analysis using the user reviews of the EC site.

Coordinated Rule Acquisition of Decision Making on Supply Chain by Exploitation-Oriented Reinforcement Learning

Product order decision-making is an important feature of inventory control in supply chains. The beer game represents a typical task in this process. Recent approaches that have applied the agent model to the beer game have shown. Q-learning performing better than genetic algorithm (GA). However, flexibly adapting to dynamic environment is difficult for these approaches because their learning algorithm assume a static environment. As exploitation-oriented reinforcement learning algorithm are robust in dynamic environments, this study, approaches the beer game using profit sharing, a typical exploitation-oriented agent learning algorithm, and verifies its results validity by comparing performances.

Evaluation of Total Quality Management Using CSR Company Reports

In recent years, serious quality accidents and quality troubles such as recalls have occurred frequently, and Total Quality Management (TQM) is important as effective management to prevent quality troubles beforehand. There is also a Quality Management Level Research [6] (TQM research) that is jointly implemented by the Union of Japanese Scientists and Engineers and the Nikkei. In the TQM survey, the rankings on the six criteria of quality management and comprehensive rankings have been announced. However, concrete TQM activities have not been announced. Meanwhile, Corporate Social Responsibility (CSR) report has published abundant descriptions about the role of customers, employees, society and management who are stakeholders of companies. In this research, we aim to evaluate and extract the characteristics of the company’s quality management activities according to the six criteria of the TQM survey using corporate CSR reports.

Visualization of online customer reviews and evaluations based on Self-organizing Map.

In recent times, it has become easier to collect large quantities of customer reviews of products and services through the Internet. Thus, text-mining has become evermore important for various businesses. However, current techniques to visualize the correspondence relation between customer reviews and evaluation information are insufficient. The purpose of this paper is to propose a new method of visualizing information using a Self-organizing Map(SOM) that is robust for text data that is non-linear and multi-collinear. Our method involves, probabilistic Latent Semantic Indexing (pLSI) which does not require weighting for the dimension contraction of a word vector. Furthermore, we also propose a method to visualize the distribution of evaluation information on SOM. In order to assign a suitable value to dead nodes and nodes without evaluation values, we redefine the interpolation formula for the evaluation value. To confirm the effectiveness and accuracy of our proposal, we use our method to visualize customer review data on a major E-Commerce website.

Extracting Important Knowledge from Multiple Markets Using Transfer Learning

The aim of this study is to extract a customer’s needs from their reviews of an electronic commerce (EC) site using transfer learning. Transfer learning involves retaining and applying the knowledge learned from one or more tasks to efficiently develop an effective hypothesis for a new task. Recently, with the spread of EC sites, customer reviews have become a beneficial information source, as they include customers’ opinions or product reputations, and can attract attention. However, this information is too huge to browse conveniently. Moreover, to develop new products with a competitive advantage, it is necessary to incorporate customers’ opinions. Therefore, it is necessary to extract the customers’ opinions from the enormous amount of customer reviews.

A Quality Table-Based Method for Sentiment Expression Word Identification in Japanese

Identifying and summarizing opinions from online reviews is a valuable and challenging task and aspect-level sentiment analysis is a research-based approach to this task. Sentiment expression word identification is important sentiment identification task since many unique expression words appear in each entity domain and it is confirmed that text data from the internet has many collateral expressions. Generally, syntax-based model is applied to sentiment expression word identification method. Syntax-based model can consider low frequency word; however, we need to consider many syntax relations and that may be not practical. Therefore, it is difficult to identify sentiment expression words with syntax-based model. This paper proposes quality table-based method for sentiment expression word identification. The method identifies sentiment expression words with supervised learning. The training set is created with both seed expression-aspect and word-aspect deployment based on characteristic of quality table’s relation. This paper proposes a non-syntax and relation-based model in order to solve syntax-based models’ problems. This paper carries out an experimental test, demonstrates how many unique SEWs are extracted, and verifies the coverage of SEW with annotated text.

An ensemble model of self-organizing maps for imputation of missing values

The purpose of this study is to improve the accuracy of missing value estimation by using self-organizing maps (SOMs). We propose an ensemble model of self-organizing maps, a new method for the imputation of missing values, which is an important preprocessing step in data analysis. Learning results of self-organizing maps have diversity because the self-organizing maps learning algorithm has a dependence on initial values; this property can be used to contribute to improving the accuracy of ensemble learning. In this study, we estimated missing values by an ensemble learning procedure that leverages the initial value dependence of the SOM. We tested the effectiveness of the proposed method by computational experiments using data published in the UCI Machine Learning Repository. Our experimental results confirmed that the proposed method produced higher accuracy than a conventional SOM when estimating values that were randomly set to missing.

The Impact of Adaptive Regularization of the Demand Predictor on a Multistage Supply Chain Simulation

The supply chain is difficult to control, which is representative of the bullwhip effect. Its behavior under the influence of the bullwhip effect is complex, and the cost and risk are increased. This study provides an application of online learning that is effective in large-scale data processing in a supply chain simulation. Because quality of solutions and agility are required in the management of the supply chain, we have adopted adaptive regularization learning. This is excellent from the viewpoint of speed and generalization of convergence and can be expected to stabilize supply chain behavior. In addition, because it is an online learning algorithm for evaluation of the bullwhip effect by computer simulation, it is easily applied to large-scale data from the viewpoint of the amount of calculation and memory size. The effectiveness of our approach was confirmed.

Knowledge Extraction About Brand Image Using Information Retrieval Method

The purpose of this study is to extract characteristic words as the information of expression pertaining to a brand’s image, from the language resources that have accumulated by users on Twitter. In this study, we analyzed Twitter data related to brands extracted the characteristic representations by using Okapi BM25, which is a ranking function that has been recently introduced in information retrieval. To confirm the validity of our approach, we conduct comparative experiments on the Twitter data of several Japanese automobile brands using BM25 and TF-IDF. By using the BM25, the extraction of keywords that are meaningful and buried in high frequency terms can be expected.