第68行: |
第68行: |
| {| class="wikitable" | | {| class="wikitable" |
| |+ | | |+ |
| + | 各种因果度量方法及其形式化公式 |
| !序号 | | !序号 |
| !名称 | | !名称 |
− | !公式 | + | !形式化公式及其与因果基元的关系 |
| !备注 | | !备注 |
| |- | | |- |
| |1 | | |1 |
− | | | + | |David Hume的恒常连结 |
− | |<math>deg=\sum_{e\in E}P(e\mid c) deg(e)=1-\frac{H(e\mid C)}{\log_{2}n}</math> | + | |<math>CS_{Galton}(e,c)=P(c)P(C\backslash c)[P(e\mid c)-P(e\mid C\backslash c)]=P(c)P(C\backslash c)[suff(e,c)+nec(e,c)-1]</math> |
| | | | | |
| |- | | |- |
| |2 | | |2 |
− | | | + | |Eells 的因果关系度量是概率提升 |
− | | | + | |<math>CS_{Eells}=P(e\mid c)-P(e\mid C\backslash c)=suff(e,c)+nec(e,c)-1</math> |
| | | | | |
| |- | | |- |
| |3 | | |3 |
− | | | + | |Suppes将因果关系度量为概率提升 |
− | | | + | |<math>CS_{Suppes}(c,e)=P(e\mid c)-P(e\mid C)=suff(e,c)-nec^{\dagger}(e)</math> |
| | | | | |
| |- | | |- |
| |4 | | |4 |
− | | | + | |程氏的因果归因 |
− | | | + | |<math>CS_{Cheng}(c,e)=\frac{P(e\mid c)-P(e\mid C\backslash c)}{1-P(e\mid C\backslash c)}=\frac{suff(e,c)+nec(e,c)-1}{nec(e,c)}</math> |
| | | | | |
| |- | | |- |
| |5 | | |5 |
− | | | + | |Lewis的反事实因果理论 |
− | | | + | |<math>CS_{Lewis}(c,e)=\frac{P(e\mid c)-P(e\mid C\backslash c)}{P(e\mid c)}=\frac{suff(e,c)+nec(e,c)-1}{suff(e,c)}</math> |
| | | | | |
| |- | | |- |
| |6 | | |6 |
− | | | + | |Judea Pearl的因果关系测量方法 |
− | | | + | |<math>\mathrm{PNS}=P(e\mid c)-P(e\mid C\backslash c)=suff(e,c)+nec(e,c)-1</math><math>\mathrm{PN}=\frac{P(e\mid c)-P(e\mid C\backslash c)}{P(e\mid c)}=\frac{suff(e,c)+nec(e,c)-1}{suff(e,c)}</math> |
− | | | + | |
| + | <math>\mathrm{PS}=\frac{P(e\mid c)-P(e\mid C\backslash c)}{1-P(e\mid C\backslash c)}=\frac{suff(e,c)+nec(e,c)-1}{nec(e,c)}</math> |
| + | |PNS对应关联层级,等价于<math>CS_{Eells}</math> |
| + | |
| + | PN对应干预层级,等价于<math>CS_{Lewis}</math> |
| + | |
| + | PS对应反事实层级,等价于<math>CS_{cheng}</math> |
| |- | | |- |
| |7 | | |7 |
− | | | + | |最接近的可能世界因果关系 |
− | | | + | |<math>CS_{Lewis CPW}=\frac{P(e\mid c)-P(e\mid\bar{c}_{CPW})}{P(e\mid c)}</math> |
− | | | + | |其中<math>\bar{c}_{CPW}=\min_{c'}D_H(c,c')</math>,<math>D_H(c,c')</math>为<math>c</math>和<math>c'</math>之间的汉明距离 |
| |- | | |- |
| |8 | | |8 |
− | | | + | |位翻转措施 |
− | | | + | |<math>CS_{bit-flip}(e,c)=\frac{1}{N}\sum_{i}^{N}\sum_{e^{\prime}\in E}P(e^{\prime}\mid c_{[i]})D_{H}(e,e^{\prime})</math> |
− | | | + | |其中<math>c_{[i]}</math>对应于第<math>i^{th}</math>位被翻转的状态(例如,如果 c = 000,则 c[3] = 001),<math>D_H(e,e')</math>为<math>e</math>和<math>e'</math>之间的汉明距离 |
| |- | | |- |
| |9 | | |9 |
− | | | + | |实际因果关系和结果信息 |
− | | | + | |<math>ei(c,e)=\log_2\frac{P(e\mid c)}{P(e\mid C)}=\log_2n[det(e,c)-deg(c)]</math> |
| | | | | |
| |- | | |- |
| |10 | | |10 |
− | | | + | |有效信息(EI) |
− | | | + | |<math>EI=\sum_{e\in E,c\in C}P(e,c)ei(c,e)=\log_{2}n[det-deg]</math> |
| | | | | |
| |} | | |} |