[论文阅读] Transductive Multilabel Learning via Label Set Propagation

#[论文阅读] Transductive Multilabel Learning via Label Set Propagation

论文下载链接：Transductive Multilabel Learning via Label Set Propagation

Abstract

• We estimate the label sets of the unlabeled instances effectively by utilizing the information from both labeled and unlabeled data.
• We first formulate the transductive multilabel learning as an optimization problem of estimating label concept compositions.

##1 Introduction

• Mutilabel learning 的重要性
• 常规方法是基于 supervised settings
• 文章考虑利用未标记数据来帮助提高 multilabel classification performance, 即 transductive learning
• Transductive learning: is a type of approaches to exploit unlabeled data in classification processes. Transductive learning assumes all the testing data are available, and the goal is to achieve better performances on these testing data by exploiting the unlabeled testing data in the classification process.
• the major research challenges on transductive multilabel learning:
  – lake of labeled data
  – multiple label: Conventional transductive learning
  approaches focus on single-labeled classification problems, strictly following the assumption that each instance has only one label
• 本文贡献：TRAM、 a closed-form solution to the optimization problem

2 Related Work

• Multilabel Classification
• Transductive Learning
  – Three main paradigms of approaches to utilize unlabeled data is, semi-supervised learning, transductive learning and active learning.
  – Two types of Transductive Learning is Transductive SVMs and graph-based methods

3 Problem Formulation

• the key issue: how to predict a set of multiple labels for each unlabeled instance based on a limited number of labeled examples and a large number of unlabeled examples

1. How to properly estimate the composition of label concepts within the label set of an unlabeled instance based upon information from both labeled instances and all the other unlabeled instances?
2. How to predict the label set for each unlabeled instance based on the estimated label concept composition with only a limited number of training examples?

• basic idea
  – first estimate then predict
  – label concept composition: equal weights, smoothness(unlabeled to labeled, unlabeled to unlabeled) 但是我不是很理解基于此说要jointly/simultaneously 为所有未标记数据find optimal solutions
• Optimization
  – 基于kNN构建权重邻接图，使用k-d树实现，但因为k-d Tree存在维度灾难，因此在构建树之前进行了多标记降维(multilabel dimensionality reduction approach, MDDM)
  – 根据kNN结果，计算获得权重矩阵
  – 基于平滑假设，k近邻重构误差最小原则，学习未标记数据的标记置信度向量
• A Closed-Form Solution
  – Cacu