Seminar《Diffusion Processes for Retrieval Revisited》

Reference

• 原文
• Oral Presentation
• Homepage&Code
• Seminar PPT
• 相关博文

一句话表述

综述了diffusion的各种方法，根据它们的共同点得出了一个通用的框架，该框架分为三个部分，Initialization；Definition of the transition matrix；Definition of the diffusion process. 分别在每个部分给出了对应的几种方法，从而将现有的大多数diffusion的方法投影到该框架中；最后根据该框架，在几个数据集上进行（遍历）测试得到了很好的结果。

细节

• 
  
• what is diffusion processes ?
  Diffusion processes on affinity graphs for capturing the intrinsic manifold structure defined by pairwise affinity matrices.
  

the manifold, defined by the provided affinity matrix, is interpreted as a weighted graph, where each element is represented by a node, and edges connect all nodes with corresponding edge weights proportional to the pairwise affinity values. Then the pairwise affinities are re-evaluated in the context of all other elements, by diffusing the similarity values through the graph. The most common diffusion processes are based on random walks, where a transition matrix defines probabilities for walking from one node to a neighboring one, which are fixed proportional to the provided affinities. By repeatedly making random walk steps on the graph, affinities are spread on the manifold, which in turn improves the obtainable retrieval scores.

Figures

Some querstions

1. 给出了框架但是没有给出求解方法—-传统的迭代方法。
2. 关于RWR与马尔科夫相似的问题—-就是一样的
3. 马尔科夫与CRF的相似性?