浅谈知识图谱的关联视角和因果视角.pdf

1、The Correlation and Causality Perspectives of Knowledge GraphPeng Cui，Yuan Meng Tsinghua UniversityAI is stepping into high-stake areas2Shifting from Performance Driven to Risk SensitiveProblems of todays ML-Explainability3Human in the loopHuman in the loopUnexplainableUnexplainableHealth Military F

2、inance IndustryMost machine learning models are black-box models4Most ML methods are developed under I.I.D hypothesisProblems of todays ML-StabilityOOD Generalization Problem5YesMaybeNoProblems of todays ML-Stability6Cancer survival rate predictionTraining DataPredictive ModelTesting DataCity Hospit

3、alUniversity HospitalHigher income,higher survival rate.City HospitalSurvival rate is not so correlated with income.Problems of todays ML-Stability7A plausible reason:CorrelationCorrelation is the very basics of machine learning.8Correlation is not explainable9Correlation is unstable10Its not the fa

4、ult of correlation,but the way we use itThree sources of correlation:CausationCausal mechanismStable and explainableConfoundingIgnoring XSpurious CorrelationSample Selection BiasConditional on SSpurious CorrelationTYTYXTYSDogGrassSample SelectionIce Cream SalesSummerDeath RateDisease ConditionICUKno

5、wledge Graph11The relationships between entities are described by edges,and the heterogeneous network is used to express the known facts and knowledge.Knowledge Graph:Correlation v.s.CausalityRepresentation and Constructionfact-causalityInferenceKnow Why-Know How-Know WhatWhat Known-What Unknown Uti

6、lityPredictionInferenceDecision12Metaknowledge13Metaknowledge is the knowledge about knowledge.Evans,J.A.;Foster,J.G.(2011-02-11).Metaknowledge.Science.Metaknowledge14Metaknowledge is the knowledge about knowledge.Evans,J.A.;Foster,J.G.(2011-02-11).Metaknowledge.Science.A B impliesCauseThe weather i