|
LOCANET** New: result analysis ** |
Provide a depth
3 causal network (oriented graph structure) around the target using
training data only, i.e., <dataname>0_train.data and <dataname>0_train.targets. |
Depth |
Desired |
|
1 |
1 |
2 |
2 |
2 |
2 |
3 |
3 |
3 |
3 |
3 |
3 |
3 |
3 |
X |
Obtained |
Relationship |
|
P |
C |
Sp |
GC |
Si |
GP |
GGP |
uud |
N |
PS |
SC |
IL |
CP |
GGC |
Other |
|
|
|
u |
d |
du |
dd |
ud |
uu |
uuu |
uud |
udd |
udu |
ddu |
duu |
dud |
ddd |
|
1 |
Parents |
u |
0 |
1 |
1 |
2 |
1 |
1 |
2 |
2 |
2 |
2 |
2 |
2 |
2 |
3 |
4 |
1 |
Children |
d |
1 |
0 |
1 |
1 |
1 |
2 |
3 |
2 |
2 |
2 |
2 |
2 |
2 |
2 |
4 |
2 |
Spouses |
du |
1 |
1 |
0 |
1 |
2 |
1 |
2 |
2 |
2 |
1 |
1 |
1 |
1 |
2 |
4 |
2 |
Gchildren |
dd |
2 |
1 |
1 |
0 |
1 |
2 |
3 |
2 |
1 |
2 |
1 |
2 |
1 |
1 |
4 |
2 |
Siblings |
ud |
1 |
1 |
2 |
1 |
0 |
1 |
2 |
1 |
1 |
1 |
2 |
2 |
1 |
2 |
4 |
2 |
Gparents |
uu |
1 |
2 |
1 |
2 |
1 |
0 |
1 |
1 |
2 |
1 |
2 |
1 |
2 |
3 |
4 |
3 |
Ggparents |
uuu |
2 |
3 |
2 |
3 |
2 |
1 |
0 |
1 |
2 |
1 |
2 |
1 |
2 |
3 |
4 |
3 |
Uncles/Aunts |
uud |
2 |
2 |
2 |
2 |
1 |
1 |
1 |
0 |
1 |
2 |
3 |
2 |
1 |
2 |
4 |
3 |
Nieces/Nephews |
udd |
2 |
2 |
2 |
1 |
1 |
2 |
2 |
1 |
0 |
1 |
2 |
3 |
2 |
1 |
4 |
3 |
Parents of siblings |
udu |
2 |
2 |
1 |
2 |
1 |
1 |
1 |
2 |
1 |
0 |
1 |
2 |
2 |
2 |
4 |
3 |
Spouses of children |
ddu |
2 |
2 |
1 |
1 |
2 |
2 |
2 |
3 |
2 |
1 |
0 |
1 |
2 |
1 |
4 |
3 |
Parents in law |
duu |
2 |
2 |
1 |
2 |
2 |
1 |
1 |
2 |
3 |
2 |
1 |
0 |
1 |
2 |
4 |
3 |
Children of spouses |
dud |
2 |
2 |
1 |
1 |
1 |
2 |
2 |
1 |
2 |
2 |
2 |
1 |
0 |
1 |
4 |
3 |
Ggchildren |
ddd |
3 |
2 |
2 |
1 |
2 |
3 |
3 |
2 |
1 |
2 |
1 |
2 |
1 |
0 |
4 |
X |
Other |
|
4 |
4 |
4 |
4 |
4 |
4 |
4 |
4 |
4 |
4 |
4 |
4 |
4 |
4 |
0 |
LUCAS |
1.09 |
|
|
|
LUCAP |
1.39 |
|
|
|
REGED |
0.42 |
0.2 |
1.01 |
0.34 |
SIDO |
3.44 |
3.46 |
3.25 |
|
CINA |
2.15 |
2.23 |
2.18 |
1.74 |
MARTI |
|
0.36 |
1.68 |
|
|
LUCAS |
LUCAP |
REGED |
SIDO |
CINA |
MARTI |
Brown |
|
|
0.27 |
3.46 |
2.23 |
0.36 |
De-Prado-Cumplido |
|
|
|
|
3.27 |
|
Dindar |
|
|
|
|
1.70 |
|
Engin |
|
|
|
3.48 |
|
|
Kirkagaclioglu |
|
|
|
|
2.16 |
|
Mwebaze |
0.91 |
1.80 |
0.22 |
3.46 |
2.32 |
|
Oguz |
|
|
|
|
1.75 |
|
Olsen |
|
|
0.52 |
|
3.31 |
0.21 |
Tillman |
|
|
0.34 |
|
1.74 |
|
Wang |
|
|
0.50 |
3.31 |
2.17 |
0.93 |
Reference A |
0.09 |
1.09 |
0.01 |
0.64 |
0.64 |
0.02 |
Reference B |
2.36 |
1.87 |
0.16 |
1.92 |
1.89 |
0.16 |
Reference C |
2.09 |
1.43 |
3.08 |
|
1.67 |
3.01 |
Reference D |
3.56 |
3.33 |
0.22 |
3.67 |
3.64 |
0.21 |
Laura Brown and Ioannis
Tsamardinos |
A
Strategy for Making Predictions Under Manipulation (JMLR W&CP, 3:35-52,
2008) |
[Abstract][Fact
Sheet] |
Catharina Olsen |
Using mutual information
to infer causal relationships |
[Fact
Sheet] |
Robert Tillman and Joseph
Ramsey |
Fan search with Bayesian scoring
for directionality applied to the LOCANET challenge |
[Fact
Sheet] |
Mario de-Prado-Cumplido |
Discovery of Causation
Direction by Machine Learning Techniques |
[Abstract][Poster] |
Ernest Mwebaze,
John A. Quinn |
Fast Committee-Based Structure
Learning |
[Preprint] |
You Zhou, Changzhang
Wang, Jianxin Yin, Zhi Geng |
Discover Local Causal Network
around a Target with a Given Depth |
[Preprint][Fact Sheet] |
age C4 E1 corr= 0.24We show in blue the variables more often called causes and in green those more ofte called effect. The age is a definite cause and this makes sense. Occupations are sometimes called causes and sometimes called effect. It is questionable whether occupation_Exec_managerial could be an effect (cited 6 times as an effect and only two times as a cause). Also very questionable is whether race_Amer_Indian_Eskimo and educationNum could be effects. On the other hand, the fact that capitalGain and capitalGain are considered effect could make sense (you can invest only if you earn a lot). Fianlly, the causal relationship of the other features (e.g. related to marital status or occupation) is not impossible.
occupation_Prof_specialty C3 E2 corr= 0.17
fnlwgt C3 E2 corr=-0.01
maritalStatus_Married_civ_spouse C3 E3 corr= 0.44
educationNum C2 E3 corr= 0.34
occupation_Other_service C3 E4 corr=-0.16
hoursPerWeek C2 E4 corr= 0.23
relationship_Unmarried C1 E4 corr=-0.14
workclass_Self_emp_not_inc C1 E4 corr= 0.02
capitalLoss C4 E7 corr= 0.14
race_Amer_Indian_Eskimo C1 E5 corr=-0.03 <--??? unrelated
maritalStatus_Divorced C1 E5 corr=-0.13
workclass_State_gov C1 E5 corr= 0.01
occupation_Exec_managerial C2 E6 corr= 0.22 <-- ?? why an effect
capitalGain C3 E8 corr= 0.22
maritalStatus_Married_civ_spouseThe majority of these highlly correlated features are found in the top half of the features preferred by the pot-luck challenge participants (outlined in red). Three (educationNum, marital_staus_Never_married, and sex) are tie with the feature just in the middle (outlined in orange) -- educationNum is cited 5 times as a direct cause or effect. The feature outline in black is just below the features with median score. So, overall, it is fair to say that correlation plays an important role in the selection of features, which are in the neighborhood if the target.
relationship_Husband
educationNum
maritalStatus_Never_married
age
hoursPerWeek
relationship_Own_child
capitalGain
sex
occupation_Exec_managerial
relationship_Not_in_family
occupation_Prof_specialty
occupation_Other_service
capitalLoss
relationship_Unmarried
Features outside the MB may be very predictive if they are hubs (like educatioNun and age). The top ranking features from Figure 1 include capitalGain, capitalLoss occupation_Exec_managerial, and maritalStatus_Married_civ_spouse, but not educationNum and age. From the point of view of causal discovery, the two variables educationNum and age may be indirectly related to the target. For instance educationNum and age may be a cause of "occupation". Since there are many possible occupations, the information is diluted among several variables, some of which may not be found significantly related to the target by the algorithms.maritalStatus_Married_civ_spouse
educationNum
capitalGain
occupation_Exec_managerial
capitalLoss
age