A Study on Construction Method for Interpretable Models based on Visualization of Multi-dimensional Nonlinear Data
| Project/Area Number |
16500126
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Sensitivity informatics/Soft computing
|
|---|
| Research Institution | Nagoya University |
|---|
Principal Investigator |
FURUHASHI Takeshi Nagoya University, Graduate School of Engineering, Professor, 大学院工学研究科, 教授 (60209187)
|
|---|
| Project Period (FY) |
2004 – 2006
|
| Project Status |
Completed (Fiscal Year 2006)
|
| Budget Amount *help |
¥3,300,000 (Direct Cost: ¥3,300,000)
Fiscal Year 2006: ¥800,000 (Direct Cost: ¥800,000)
Fiscal Year 2005: ¥800,000 (Direct Cost: ¥800,000)
Fiscal Year 2004: ¥1,700,000 (Direct Cost: ¥1,700,000)
|
|---|
| Keywords | Data visualization / Interpretability of clusters / Clustering / Clustering of interpretable models / Kansei data analysis / Individuality / SD Evaluation / Impression word / モデルの明示性 / モデリング / FCM / EMアルゴリズム / カーネル関数 / 個人性の可視化 |
|---|
| Research Abstract |
This study was aimed at establishing a visualization methodology for interpretation of multi-dimensional nonlinear data. Knowledge extraction from questionnaire data, for example, can be utilized for planning of new products. However, it was difficult to find a statistically significant result from a small number of multi-variate data. It is expected that a support for the knowledge extraction can be carried out by visualizing the data. This study achieved the following two major results:
1. The criteria for the interpretability of visualized clusters were defined, and a clustering method and a dimension reduction method based on the criteria were developed.
The clustering method and the dimension reduction method were based on the unique criteria, i.e. the interpretability of visualized clusters. These methods generated separated clusters in the visualized space. Experiments using benchmark problems such as iris data and wine data demonstrated that interpretable clusters were obtained b
… More
y the proposed methods.
2. A clustering method and a visualization method based on the characteristics contained in subjects' evaluation values were developed.
The clustering method was a unique one based on the similarity of distances between questionnaire objects. This clustering method applied Orthogonal Procrustes Analysis (OPA) to cluster subjects who had similar evaluation distances between objects. The OPA identified clusters without depending on differences in subjects' understanding of questions.
The visualization method was a new multi-dimensional scaling method that visualized asymmetrical distances between questions based on correlations of evaluation values of two subjects. This method enabled analysis on differences in subjects' understanding of questions. The differences could be that a question was understood differently by the two subjects, or that an understanding of a question by one subject was the same as that of another question by the other subject.
The visualization method was extended to visualize differences in understanding between clusters. The distance between questions was defined based on the correlation of evaluation values of subjects in two clusters. This method provided a new view point for the analysis of characteristics of obtained clusters. Less
|
Report
(4 results)
Research Products
(22 results)
