当前位置:网站首页>论文笔记: 多标签学习 BP-MLL
论文笔记: 多标签学习 BP-MLL
2022-06-13 07:21:00 【闵帆】
摘要: 分享对论文的理解. 原文见 Zhang, M.-L., & Zhou, Z.-H. (2006). Multi-label neural networks with applications to functional genomics and text categorization. IEEE transactions on Knowledge and Data Engineering, 18, 1338–1351.
1. 论文贡献
The first multilabel neural network algorithm.
多标签学习的第一个神经网络算法.
2. 主要思想
The labels belonging to an instance should be ranked higher than those not belonging to that instance.
例: 如果图片里面有猫但没狗, 则猫对应的预测值应该比狗的大.
更实际的例子: 对猫这个标签的预测值为 0.6, 对狗为 0.3, 则惩罚较小 (机器学习里面一般惩罚值都大于 0); 否则惩罚很大.
进一步解读: 成对的比较.
3. 符号系统
| 符号 | 含义 | 说明 |
|---|---|---|
| X = R d \mathcal{X} = \mathbb{R}^d X=Rd | d d d 维特征空间 | |
| Y = { 1 , 2 , … , Q } \mathcal{Y} = \{1, 2, \dots, Q\} Y={ 1,2,…,Q} | 标签空间 | 使用整数集合而不是 { − 1 , 1 } d \{-1, 1\}^d { −1,1}d |
| h : X → 2 Y h: \mathcal{X} \to 2^\mathcal{Y} h:X→2Y | 分类器 | 使用幂集 |
| f : X × Y → R f: \mathcal{X} \times \mathcal{Y} \to \mathbb{R} f:X×Y→R | 回归器 | 为每个对象的每个标签预测一个概率 |
说明:
- 从回归器到分类器, 只需要增加一个阈值.
- 有些 (更为流行的) MLL 评价指标并不需要输出 + 1 +1 +1 或 − 1 -1 −1, 而是标签的排序.
4. 算法核心

E = ∑ i = 1 m E i = ∑ i = 1 m 1 ∣ Y i ∣ ∣ Y i ‾ ∣ ∑ ( k , l ) ∈ Y i × Y i ‾ exp ( − ( c k i − c l i ) ) , (1) E = \sum_{i=1}^m E_i = \sum_{i=1}^m \frac{1}{\vert Y_i \vert \vert \overline{Y_i} \vert} \sum_{(k, l) \in Y_i \times \overline{Y_i}} \exp(-(c_k^i - c_l^i)), \tag{1} E=i=1∑mEi=i=1∑m∣Yi∣∣Yi∣1(k,l)∈Yi×Yi∑exp(−(cki−cli)),(1)
其中:
- Y i ‾ = Y ∖ Y i \overline{Y_i} = \mathcal{Y} \setminus Y_i Yi=Y∖Yi, 即 Y i Y_i Yi 的补集.
- c k i = f ( x i , k ) c_k^i = f(\mathbf{x_i}, k) cki=f(xi,k) 是神经网络对于 x i \mathbf{x_i} xi 第 k k k 个标签的预测值.
- 第一个求和是针对所有对象.
- 第二个求和本质上是双重循环, 如 ∣ Y i ∣ = 10 \vert Y_i \vert = 10 ∣Yi∣=10, Q = 100 Q = 100 Q=100, 则需要求 10 × 90 = 900 10 \times 90 = 900 10×90=900 对.
- 当 exp \exp exp 的参数小于 0 时, 损失较小, 否则相当大.
5. 小结
- 成对比较的思想.
- 一篇机器学习的论文, 一个损失函数就够了.
- 可以把 exp \exp exp 替换成另外的单增函数, 或者改变网络结构, 以此来进行小的创新.
边栏推荐
- The management practice of leading enterprises has proved that what is the core of sustainable development of enterprises?
- redis-3. Redis list, set, hash, sorted_ set、skiplist
- Can flush open a stock account? Is it safe?
- FSM状态机
- Local file upload FTP or remote directory
- TXT_ File encryption and compression
- redis-0. Introduction to redis and NiO principle (random talk)
- 全志V3S环境编译开发流程
- Interview questions must be asked - Optimization of large table Pagination
- 汇编语言基础:寄存器和寻址方式
猜你喜欢

redis-2. Redis string type & bitmap

redis-5. Redis' RDB, fork, copyonwrite, AOF, RDB & AOF are mixed

Xuanwu cloud technology passed the listing hearing: the performance fluctuated significantly, and chenyonghui and other three were the controlling shareholders

How worker threads in the thread pool are recycled

Tidb dashboard modify port

Powerdispatcher reverse generation of Oracle data model

A solution to the problem that there is always a newline character when C merges multiple RichTextBox contents

号称下一代监控系统 来看看它有多牛逼

基于FPGA的ds18b20温度传感器使用

量化框架backtrader之一文讀懂Analyzer分析器
随机推荐
Simple understanding of basic language of C language
One article of quantitative framework backtrader read analyzer
部署RDS服务
Application of DS18B20 temperature sensor based on FPGA
C Advanced Programming - features
C # related knowledge points
关于#数据库#的问题:PGADMIN4 编辑sql窗口问题
SDN basic overview
Relevant knowledge under WinForm
redis-5. Redis' RDB, fork, copyonwrite, AOF, RDB & AOF are mixed
An example of CSRF attack defense in web application scenarios
How worker threads in the thread pool are recycled
号称下一代监控系统 来看看它有多牛逼
Functions about Oracle.
Detailed description of drawing ridge plot, overlapping densities of overlapping kernel density estimation curve, facetgrid object and function sns Kdeplot, function facetgrid map
What languages can be decompiled
基于ESP32CAM实现WebSocket服务器实时点灯
Ticdc introduction
DM Experiment 6: filter replication
Issues related to C # delegation and events