Causal Analysis

Cyria7 Lv2

文献阅读记录

CausalVAE

论文地址:https://arxiv.org/abs/2004.08697

代码地址:https://github.com/huawei-noah/trustworthyAI/tree/master/research/CausalVAE

  1. 论文主要内容

华为诺亚方舟实验室提出基于因果结构模型的解耦表征学习方法。该方法挑战了传统解耦任务的独立性假设,第一次采用因果结构假设物理世界的概念关系作为解耦表征的依据。并从理论上证明了采用该方法得到的解耦出的概念表征的可识别性。作者在两组合成数据及人脸数据CelebA上共测试了4组不同类型因果结构下的解耦效果,实验证明该方法不仅可以分辨出概念表征,并可以发现概念之间的因果结构。

评价指标

MIC(Maximal Information Coefficient)

MIC(最大信息系数),是用于衡量两个变量X和Y之间的关联程度,线性或非线性的强度,常用于机器学习的特征选择。

详细计算方法参考:https://blog.csdn.net/FontThrone/article/details/85227239

TIC(Total Information Coefficient)

MIC指标是通过不同的网格划分,计算不同划分状态下的某一个最大信息系数,而TIC则是将不同的划分的互信息系数相加得到

参考资料

因果结构学习:
DAG-GNN: DAG structure learning with graph neural networks
https://github.com/fishmoon1234/DAG-GNN

A Graph Autoencoder Approach to Causal Structure Learning
https://github.com/huawei-noah/trustworthyAI/tree/master/gcastle/castle/algorithms/gradient/gae

Masked Gradient-Based Causal Structure Learning
https://github.com/huawei-noah/trustworthyAI/tree/master/gcastle/castle/algorithms/gradient/mcsl

解耦表示学习:
各种VAE的实现:https://github.com/AntixK/PyTorch-VAE

Interactive Disentanglement: Learning Concepts by Interacting with their Prototype Representations
https://github.com/ml-research/xiconceptlearning

因果表示学习:
CausalVAE: Structured Causal Disentanglement in Variational Autoencoder
https://github.com/huawei-noah/trustworthyAI/tree/master/research/CausalVAE

SCM-VAE: Learning Identifiable Causal Representations via Structural Knowledge
https://github.com/Akomand/SCM-VAE

Shadow Datasets, New challenging datasets for Causal Representation Learning(主要是因果表示学习的数据集)
https://github.com/Jiagengzhu/Shadow-dataset-for-crl

  • Title: Causal Analysis
  • Author: Cyria7
  • Created at : 2023-09-06 14:17:59
  • Updated at : 2023-10-18 11:58:10
  • Link: https://cyria7.github.io/2023/09/06/causal/
  • License: This work is licensed under CC BY-NC-SA 4.0.
Comments