Machine Learning based Inference on Causality in Graphs
- 주제(키워드) Machine Learning , Causality , Causal Inference , Graph , Neural Network
- 발행기관 아주대학교
- 지도교수 신현정
- 발행년도 2021
- 학위수여년월 2021. 2
- 학위명 박사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000030879
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Causalities between data points are represented as directions in a graph. It can play an important role in explanation, prediction, interpretation, and decision making. Also, since several studies have shown that the algorithms are improved when using the directed graph, causal inference between data points may contribute to improving the prediction performance of the algorithms. The purpose of this study is to infer causalities with a high possibility through learning and extracting from data among a huge number of possible causalities. The proposed methods consist of two approaches: (a) machine learning- and (b) data mining-based causal inference on graphs. For the machine learning model, ELFNet, an Edge Labeling and Feature inference neural network is proposed. It predicts the directionality between nodes with edge labeling and infers the edge features simultaneously. For the data mining model, αDCFC, extracts causality between elements from text data and quantify the degree of the causal relation is proposed. The proposed methods provide insight to understand how data points are related to each other. Causal inference also can be used in a variety of analyzes by using the results. Therefore, applications that can be applied to a disease network when causal information is provided are presented in two aspects: (a) the shortest path search algorithm for finding causal disease chains and (b) the machine learning algorithm for disease comorbidity scoring. The results of applications demonstrate that causal information provides extended insight or improving algorithms.
more목차
1. Introduction 1
2. Causal Inference on Graphs 7
2.1. Machine Learning-based Causal Inference 9
2.1.1. ELFNet: Edge Labeling and Feature Inference Neural Network 10
2.1.2. Experiments: Predicting Directionality with Edge Feature Inference 16
2.2. Data Mining-based Causal Inference 26
2.2.1. Causality Extraction Method from Text 29
2.2.2. Experiments: Causal Disease Network Construction 35
3. Applications of Causal Graphs 42
3.1. Inference on Causal Disease Chain 44
3.1.1. Finding Causal Disease Chains from Disease Networks 45
3.1.2. Experiments 58
3.2. Disease Comorbidity Scoring with Causal Disease Networks 69
3.2.1. Semi-Supervised Scoring for Causal Graphs 70
3.2.2. Experiments 75
4. Conclusions 87
References 92
