| Project/Area Number |
09480065
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (B)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Intelligent informatics
|
|---|
| Research Institution | Osaka University |
|---|
Principal Investigator |
MOTODA Hiroshi ISIR, Osaka University, Professor, 産業科学研究所, 教授 (00283804)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
HORIVCHI Tadashi ISIR, Osaka University, Research Associate, 産業科学研究所, 助手 (50294129)
WASHIO Takashi ISIR, Osaka University, Associate Professor, 産業科学研究所, 助教授 (00192815)
MIZOGVCH Riichiro ISIR, Osaka University, Professor, 産業科学研究所, 教授 (20116106)
|
|---|
| Project Period (FY) |
1997 – 1999
|
| Project Status |
Completed (Fiscal Year 1999)
|
| Budget Amount *help |
¥12,300,000 (Direct Cost: ¥12,300,000)
Fiscal Year 1999: ¥1,100,000 (Direct Cost: ¥1,100,000)
Fiscal Year 1998: ¥4,700,000 (Direct Cost: ¥4,700,000)
Fiscal Year 1997: ¥6,500,000 (Direct Cost: ¥6,500,000)
|
|---|
| Keywords | Machine Learning / Graph-Based Induction / Inductive Inference / Classification Rule Learning / Command Prediction / User Interface / 帰納推理 |
|---|
| Research Abstract |
Computer needs to come closer to humans if it is to be a good partner of its user as an easy-to-use tool by interpreting its user's intention, learning her preference and responding in accordance to her expertise of computer usage. This research targeted to develop a user-adaptive interface that learns user's preference from her past usage and predicts the next command. The key factor is to devise a right kind of machine learning technique that best suits to this needs. It was recognized that narrowing down the context in which the user was working is crucial, and to do this, use of the tree structured data that involve both command sequence data and process I/O data was essential. An efficient algorithm based on the notion of pairwise chunking was developed which enabled to induce a classifier in real time from a tree structured data. Evaluation results using both artificial and real data show that the prediction is accurate enough, and the implemented interface gradually improves its performance as it learns its user's preference and comes to respond differently for a different user. Further, the learning part was made an independent program and expanded to handle general graph structured data, i.e., directed/undirected graph that has colored/uncolored nodes and links with/out loops (including self-loops). Its computational complexity is confirmed to be linear to the size of graph. It was applied to finding typical patterns from WWW browsing history data provided by a commercial provider and to discovering characteristic substructures that are typical to carcinogen of organic chlorides. Both experiments gave satisfactory results.
|
