| 研究課題/領域番号 |
22K12150
|
|---|
| 研究種目 |
基盤研究(C)
|
| 配分区分 | 基金 |
|---|
| 応募区分 | 一般 |
|---|
| 審査区分 |
小区分61030:知能情報学関連
|
|---|
| 研究機関 | 京都大学 |
|---|
研究代表者 |
|
|---|
| 研究期間 (年度) |
2022-04-01 – 2025-03-31
|
| 研究課題ステータス |
交付 (2023年度)
|
| 配分額 *注記 |
4,160千円 (直接経費: 3,200千円、間接経費: 960千円)
2024年度: 1,300千円 (直接経費: 1,000千円、間接経費: 300千円)
2023年度: 1,300千円 (直接経費: 1,000千円、間接経費: 300千円)
2022年度: 1,560千円 (直接経費: 1,200千円、間接経費: 360千円)
|
|---|
| キーワード | Graph neural networks / Convex Clustering / machine learning / Machine learning / Structured data / Deep learning / Sparse learning |
|---|
| 研究開始時の研究の概要 |
There are three directions of this research project:
(1) investigating original machine learning models for complicated structures,
(2) designing novel structure discovery tools incorporating domain knowledge, and
(3) discovering new biomedical knowledge to be used by domain experts.
|
| 研究実績の概要 |
In this year, we are working on representation of data that are faithful to the original features as well as having cluster structures. We investigated the method of convex clustering to obtain a representation using a convex program, which is efficient and globally optimal.
The key idea is to assume that data follows cluster structures. For that, we cluster the data using convex clustering. The advantage of convex clustering is that it is a convex program that guarantees optimality. Another advantage is that it offers a relaxation of k-means and agglomerative clustering algorithms, offering potential advantages of the two algorithms.
Our main work here is to analyze analytically what are the clusters that are obtained by convex clustering, pros and cons compared to the other two algorithms. We found that convex cluster only can learn convex clusters. This is similar to k-means and different from agglomerative clustering. We also found that the clusters can be bounded in balls, making them round-shaped. These clusters are found to have gaps between them. These properties show that convex clustering found rather specific types of clusters, rather inflexible compare to the other algorithms.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
3: やや遅れている
理由
We are working on a particular problem with the difficulty of understanding the formulation of convex clustering, which has not been well studied before.
|
| 今後の研究の推進方策 |
We plan to continue working on finding suitable representations of data from original features with additional information such as graphs that are guaranteed to extract more information compared to currently used methods.
|
