/Multi-label-research

paper reading notes & code

Multi-label Research

Category

  1. 按文章目标（mainly based on The Emerging Trends of Multi-Label Learning）
  • 综述类
Title Description State Repository
A Review On Multi-Label Learning Algorithms 周志华老师的，更多地关注于机器学习方法，8种常见模型 ✔️ None
The Emerging Trends of Multi-Label Learning Anchor，image-20221104210000301 ✔️ None

  • XMLC(Extreme multi-label learning)

    ⭐ one vs all.

    传统的就是针对每一个标签训练一个分类器，削弱了每个标签之间的关联性。

    可推荐阅读：

Title Description State Repository
DiSMEC - Distributed Sparse Machines for Extreme Multi-label Classification ❌ None
Deep Extreme Multi-label Learning ❌
*ML-Decoder: Scalable and Versatile Classification Head 可以面向特别多标签的情况进行处理

​ :star: Tree

Title Description State Repository
Extreme F-measure maximization using sparse probability estimates ❌ None
A no-regret generalization of hierarchical softmax to extreme multi-label classification ❌ None
Fastxml: a fast, accurate and stable tree-classifier for extreme multi-label learning classic ❌ None

​ :star: Embedding

Title Description State Repository
Sparse local embeddings for extreme multi-label classification SLEEC，经典方法！ ❌ None
Partial Multi-Label Learning by Low-Rank and Sparse Decomposition. ❌ None
Query2Label: A Simple Transformer Way to Multi-Label Classification Transformer的decoder学习label embedding，思路其实很简单，用一个Embedding层把它和backbone输出的channel数相match即可，backbone输出[2048,14,14]的feature map，2048是channel, 也是embedding，所以label的embedding大小也是2048。 ✔️ https://github.com/SlongLiu/query2labels
  • PML（Partial multi-label learning）
Title Description State Repository
Partial Multi-Label Learning classic ❌ None
Partial Multi-Label Learning with Label Distribution ❌ None
Partial Multi-Label Learning via Credible Label Elicitation ❌ None

  • SSMLC（semi-supervised MLC）

  • MLML（Multi-label learning with missing label）

Title Description State Repository
Asymmetric Loss For Multi-Label Classification classic，改loss function，提出ASL，对正负标签施加不同的focal weight ✔️ Alibaba-MIIL/ASL: Official Pytorch Implementation of: "Asymmetric Loss For Multi-Label Classification"(ICCV, 2021) paper (github.com)
Simple and Robust Loss Design for Multi-Label Learning with Missing Labels 改loss function，加了Hill Loss, SPLC和focal margin ✔️ xinyu1205/robust-loss-mlml (github.com)
Learning a Deep ConvNet for Multi-label Classification with Partial Labels 多篇文章提到！ ❌ None
  1. 按多标签分类中标签数量分类
  • single label
Title Description State Repository
Multi-Label Learning from Single Positive Labels 改loss function ✔️
Acknowledging the Unknown for Multi-label Learning with Single Positive Labels ❌
  1. 按文章的改进方向
  • improve loss function

(注：标*的代表上面已有相同的item)

Title Description State Repository
*Multi-Label Learning from Single Positive Labels 改loss function ✔️
*Simple and Robust Loss Design for Multi-Label Learning with Missing Labels 改loss function，加了Hill Loss, SPLC和focal margin ✔️
*Asymmetric Loss For Multi-Label Classification classic，改loss function，提出ASL，对正负标签施加不同的focal weight ✔️
  • model label correlation ( graph network )
Title Description State Repository
Multi-Label Image Recognition with Joint Class-Aware Map Disentangling and Label Correlation Embedding ❌
Multi-Label Image Recognition With Graph Convolutional Networks ❌
  • change model structure( backbone| head )
Title Description State Repository
TResNet: High Performance GPU-Dedicated Architecture 改loss function ❌
*Query2Label: A Simple Transformer Way to Multi-Label Classification attention-based head，主要使用的是Transformer 的decoder部分，学习的是label embedding ✔️
ML-Decoder: Scalable and Versatile Classification Head ✔️ https://github.com/Alibaba-MIIL/ML_Decoder

OOD:

感觉和zero-shot learning 有点关系，因为zero-shot是从已知类中学习，预测未知类。

This is usually done by sharing knowledge between the seen classes (that were used for training) and the unseen classes via a text model.