当前位置:网站首页>论文笔记: 多标签学习 DM2L
论文笔记: 多标签学习 DM2L
2022-06-24 06:56:00 【闵帆】
摘要: 分享对论文的理解. 原文见 Ma, Z.-C., & Chen, S.-C. (2021). Expand globally, shrink locally: Discrimi-nant multi-label learning with missing labels. Pattern Recognition, 111, 107675.
1. 论文贡献
- 从全局和局部两个方面同时优化;
- 用核函数支撑非线性变换;
- 理论分析到位.
2. 基本符号
| 符号 | 含义 | 说明 |
|---|---|---|
| X ∈ R n × d \mathbf{X} \in \mathbb{R}^{n \times d} X∈Rn×d | 属性矩阵 | |
| X k ∈ R n k × d \mathbf{X}_k \in \mathbb{R}^{n_k \times d} Xk∈Rnk×d | 具有第 k k k 个标签的属性子矩阵 | |
| Y ∈ { − 1 , + 1 } n × c \mathbf{Y} \in \{-1, +1\}^{n \times c} Y∈{ −1,+1}n×c | 标签矩阵 | |
| Y ~ ∈ { − 1 , + 1 } n × l \tilde{\mathbf{Y}} \in \{-1, +1\}^{n \times l} Y~∈{ −1,+1}n×l | 观测到的标签矩阵 | |
| Ω = { 1 , … , n } × { 1 , … , c } \mathbf{\Omega} = \{1, \dots, n\} \times \{1, \dots, c\} Ω={ 1,…,n}×{ 1,…,c} | 观测标签位置集合 | |
| W ∈ R m × l \mathbf{W} \in \mathbb{R}^{m \times l} W∈Rm×l | 系数矩阵 | 仍然是线性模型 |
| w i ∈ R m \mathbf{w}_i \in \mathbb{R}^m wi∈Rm | 某一标签的系数向量 | |
| C ∈ R l × l \mathbf{C} \in \mathbb{R}^{l \times l} C∈Rl×l | 标签相关性矩阵 | 成对相关性, 不满足对称性 |
3. 算法
基本优化目标:
min 1 2 ∥ R Ω ( X W ) − Y ~ ∥ F 2 + λ d ∥ X W ∥ ∗ (1) \min \frac{1}{2} \|R_{\Omega}(\mathbf{XW}) - \tilde{\mathbf{Y}}\|_F^2 + \lambda_d \|\mathbf{XW}\|_*\tag{1} min21∥RΩ(XW)−Y~∥F2+λd∥XW∥∗(1)
其中,
- 损失函数部分不考虑缺失值, 这个属于常规操作.
- 核正则 (nuclear norm) 部分考虑了预测的矩阵, 而不仅仅是 X W \mathbf{XW} XW, 有点奇怪.
考虑标签结构后的优化目标:
min 1 2 ∥ R Ω ( X W ) − Y ~ ∥ F 2 + λ d ( ∑ k = 1 c ∥ X k W ∥ ∗ − ∥ X W ∥ ∗ ) , (2) \min \frac{1}{2} \|R_{\Omega}(\mathbf{XW}) - \tilde{\mathbf{Y}}\|_F^2 + \lambda_d \left(\sum_{k = 1}^c \|\mathbf{X}_k\mathbf{W}\|_* - \|\mathbf{XW}\|_*\right), \tag{2} min21∥RΩ(XW)−Y~∥F2+λd(k=1∑c∥XkW∥∗−∥XW∥∗),(2)
其中,
- ∥ X k W ∥ ∗ \|\mathbf{X}_k\mathbf{W}\|_* ∥XkW∥∗ 表达了局部标签结构, 轶越小越好;
- ∥ X W ∥ ∗ \|\mathbf{XW}\|_* ∥XW∥∗ 表达了全局标签结构, 轶越大越好 (可分性更强, 信息量越高).
- 这两点就是题目的来源.
增加非线性的优化目标:
min 1 2 ∥ R Ω ( X W ) − Y ~ ∥ F 2 + λ d ( ∑ k = 1 c ∥ f ( X k ) W ∥ ∗ − ∥ f ( X ) W ∥ ∗ ) , (5) \min \frac{1}{2} \|R_{\Omega}(\mathbf{XW}) - \tilde{\mathbf{Y}}\|_F^2 + \lambda_d \left(\sum_{k = 1}^c \|f(\mathbf{X}_k)\mathbf{W}\|_* - \|f(\mathbf{X})\mathbf{W}\|_*\right), \tag{5} min21∥RΩ(XW)−Y~∥F2+λd(k=1∑c∥f(Xk)W∥∗−∥f(X)W∥∗),(5)
其中 f ( ⋅ ) f(\cdot) f(⋅) 为核函数导致的非线性变换.
4. 小结
- 又是一堆理论证明.
边栏推荐
- Vulnhub靶机:BOREDHACKERBLOG: SOCIAL NETWORK
- Blue Bridge Cup_ Queen n problem
- Industrial computer anti cracking
- [data update] Xunwei comprehensively upgraded NPU development data based on 3568 development board
- Solve the problem of notebook keyboard disabling failure
- 【点云数据集介绍】
- How does dating software cut your leeks
- MySQL source and target table row count check
- Chapter 1 overview of canvas
- Gossip: what happened to 3aC?
猜你喜欢

从 jsonpath 和 xpath 到 SPL

一文理解同步FIFO

自动化测试的未来趋势

Echart 心得 (一): 有关Y轴yAxis属性

How to use the virtual clock of FPGA?

Installation and use of selenium IDE

Swift Extension ChainLayout(UI的链式布局)(源码)

Leetcode 207: course schedule (topological sorting determines whether the loop is formed)

More appropriate development mode under epidemic situation

软件工程导论——第二章——可行性研究
随机推荐
These dependencies were not found: * core JS / modules / es6 array. Fill in XXX
Swift Extension ChainLayout(UI的链式布局)(源码)
Easyplayerpro win configuration full screen mode can not be full screen why
Ad-gcl:advantageous graph augmentation to improve graph contractual learning
3D数学基础[十七] 平方反比定理
Echart 心得 (一): 有关Y轴yAxis属性
All you know is the test pyramid?
JDBC 在性能测试中的应用
51 single chip microcomputer_ External interrupt and timer / Counter interrupt
Open cooperation and win-win future | Fuxin Kunpeng joins Jinlan organization
对于flex:1的详细解释,flex:1
VR is destined to reappear in the Jianghu?
Latest news of awtk: new usage of grid control
Optimization and practice of Tencent cloud EMR for cloud native containerization based on yarn
GraphMAE----论文快速阅读
LINQ 查询(2)
Review of postgraduate English final exam
1279_VMWare Player安装VMWare Tools时VSock安装失败解决
OC extension detects whether an app is installed on the mobile phone (source code)
【点云数据集介绍】