I am struggling to make sense of the data. The description says:
\r\n"Consider for instance a target variable B, like occurence of "lung cancer" in patients. The goal would be to find whether a factor A, like "smoking", might cause B. The objective of the challenge is to rank pairs of variables {A,\r\n B} to prioritize experimental verifications of the conjecture that A causes B."
\r\nAnd the datapoints have the following format:
\r\n| SampleID | \r\nA | \r\nB | \r\n
| train1 | \r\n2092 1143 390 1424 | \r\n5651 4449 4012 6124 | \r\n
Does this mean smoking(A) has the features: 2092 1143 390 1424 and lung cancer(B) has the features: 5651 4449 4012 6124? And if so, what excactly would it mean that smoking has feature 1 value 2092 and lung cancer feature 1 value 5651.
\r\nAny help is appreciated.
", "messages": [ { "id": "21892", "postDate": "03/31/2013 11:45:24", "content": "I am struggling to make sense of the data. The description says:
\r\n"Consider for instance a target variable B, like occurence of "lung cancer" in patients. The goal would be to find whether a factor A, like "smoking", might cause B. The objective of the challenge is to rank pairs of variables {A,\r\n B} to prioritize experimental verifications of the conjecture that A causes B."
\r\nAnd the datapoints have the following format:
\r\n| SampleID | \r\nA | \r\nB | \r\n
| train1 | \r\n2092 1143 390 1424 | \r\n5651 4449 4012 6124 | \r\n
Does this mean smoking(A) has the features: 2092 1143 390 1424 and lung cancer(B) has the features: 5651 4449 4012 6124? And if so, what excactly would it mean that smoking has feature 1 value 2092 and lung cancer feature 1 value 5651.
\r\nAny help is appreciated.
", "rawMarkdown": "", "votes": null }, { "id": "21898", "postDate": "03/31/2013 14:41:41", "content": "No, this means smoking=A, lung cancer=B (only 2 features). You get data points in a 2-dimensional space {2092, 5651}, {1143, 4449}, {390, 4012}, etc. That is you can make a plot of B vs A. The question is: Is B a cause of A or A a cause of B.
\r\n- A is a cause of B means here that B=f(A, noise) [denoted A -> B]
\r\n- A is a cause of A means A=f(B, noise) [denoted A <- B]
\r\nBut a dependency between A and B can also be explained by a common cause Z:
\r\nA=f(Z, noise) and B=f(Z, noise) [denoted A - B]
\r\nAlso, some pairs of variables are intependent (denoted A | B) but because of the limited number of data points, they may coincidentally look dependent.
\r\nSee Slide presentation: overview of the challenge. There are a few examples there.
\r\nThere is a bibliography at http://www.causality.inf.ethz.ch/cause-effect.php?page=help.
", "rawMarkdown": "", "votes": null }, { "id": "21903", "postDate": "03/31/2013 16:47:16", "content": "[quote=Isabelle;21898]
\r\nSee Slide presentation: overview of the challenge. There are a few examples there.
\r\n[/quote]
\r\nThank you the scatter plot illustrations really helps a lot in understanding the data.
", "rawMarkdown": "", "votes": null }, { "id": "21925", "postDate": "04/01/2013 04:38:28", "content": "Isabelle, Is that a typo at the beginning of
\r\n- A is a cause of A means A=f(B, noise) [denoted A <- B]
\r\nin your response to LoveSaves?
", "rawMarkdown": "", "votes": null }, { "id": "21934", "postDate": "04/01/2013 15:06:36", "content": "Yes, sorry, I meant B is a cause of A:
\r\n- B is a cause of A means A=f(B, noise) [denoted A <- B]
", "rawMarkdown": "", "votes": null }, { "id": "21969", "postDate": "04/02/2013 00:56:10", "content": "Your feed-back prompted ne to add a quiz on the first page, thanks.
\r\nI also provided some references at: http://www.causality.inf.ethz.ch/cause-effect.php?page=help
", "rawMarkdown": "", "votes": null }, { "id": "21981", "postDate": "04/02/2013 07:17:39", "content": "Thank you for the Slide Presentation.
\r\nIt seems like there are a few mix-ups on the explanations on the right of the scatter plot examples as well?
\r\nE.g. Slide 12, A=Ages, B=Wages should be A->B not A<-B?
", "rawMarkdown": "", "votes": null }, { "id": "21989", "postDate": "04/02/2013 11:56:29", "content": "Bit confused as well. Doesn't Altitude determine temperature A <- B in the slide, rather than temperature determines altitude? Also, as with the commenter above, doesn't Age determine wages? so A -> B
\r\n\r\n
Both of these slides are opposite to what is expected?
", "rawMarkdown": "", "votes": null }, { "id": "21990", "postDate": "04/02/2013 12:21:34", "content": "even bit confused !!
\r\nI am using IGCI as the causal solver, as shown in the example of IGCI,
\r\nx = randn(100,1); y = sin(x); res = igci(x,y,2,1); disp(res); % with res>0
\r\nwhile, what about asin function? it can also be interpretate as x->y,
\r\nx = randn(100,1); y = asin(x); res = igci(x,y,2,1); disp(res); % with res<0
\r\nso what's the concrete definition of 'cause' ? can we assume that the inverse operator like\r\nasin is the inverse of causal?
", "rawMarkdown": "", "votes": null }, { "id": "22005", "postDate": "04/02/2013 17:13:48", "content": "[quote=Ling Cheng;21981]
\r\nThank you for the Slide Presentation.
\r\nIt seems like there are a few mix-ups on the explanations on the right of the scatter plot examples as well?
\r\nE.g. Slide 12, A=Ages, B=Wages should be A->B not A
\r\n[/quote]
\r\n\r\n
[quote=Domcastro;21989]
\r\nBit confused as well. Doesn't Altitude determine temperature AB
\r\n\r\n
Both of these slides are opposite to what is expected?
\r\n[/quote]
\r\nThank you for pointing this out, there is an error in the slides that I rectified:
\r\nB=altitude -> A=temperature
\r\nA=age -> B=wages
\r\nSorry about that.
\r\nhttps://docs.google.com/a/chalearn.org/file/d/0B08QS7nJpK7mNUxudDdKY0MxRlE/edit
", "rawMarkdown": "", "votes": null }, { "id": "22007", "postDate": "04/02/2013 17:57:07", "content": "Thanks Isabelle!
\r\nBtw, I don't think we have access to the google docs link you posted.
", "rawMarkdown": "", "votes": null }, { "id": "22009", "postDate": "04/02/2013 18:08:12", "content": "[quote=liubenyuan;21990]
\r\neven bit confused !!
\r\nI am using IGCI as the causal solver, as shown in the example of IGCI,
\r\nx = randn(100,1); y = sin(x); res = igci(x,y,2,1); disp(res); % with res>0
\r\nwhile, what about asin function? it can also be interpretate as x->y,
\r\nx = randn(100,1); y = asin(x); res = igci(x,y,2,1); disp(res); % with res
\r\nso what's the concrete definition of 'cause' ? can we assume that the inverse operator like\r\nasin is the inverse of causal?
\r\n[/quote]
\r\nYou are touching at the heart of the problem. But there are several issues here:
\r\n1) Definition of causality. For artificial data, if B was generated as a function of A and some other "noise" variables, then A is a cause of B, by definition. For real data, we have to assume that there is\r\n an underlying data generating process that is equivalent to a mathematical function achieving the same thing. In the dataset we provided, we determined ground truth with the knowledge we have of the variables (if there is a statisticall dependency between\r\n temperature and altitude of cities, we will say that altitude causes temperature because of what we know of what these variables mean and how they have been measured).
\r\n2) Identifiability.In either case (artificial or natural), it does not mean that the causal direction is identifiable with observational data. For instance, if A and the noise are Gaussian distributed and B\r\n = w * A + noise, then it is impossible to determine whether A -> B or A
\r\n3) Experimentation: the gold standard. The gold standard for determining causal relationships is experimentations: kick A and see what happens to B. This is different from just observing values. For instance,\r\n if there is a device with a needle whose position correlates with temperature, if it is a themometer, forcing the needle to a given position will not change temparature; if it is a thermostat, it will. Our end goal is actually to find factors on which we can\r\n act to trigger desired effects. We do not expect that we will always be able to do so by just observing without experimenting. In non identifiable cases, we will have to resort to experimenting (e.g. setting A to given values and measure B while randomizing\r\n over other factors and vice versa). For instance, in the smoking and lung cancer example, we would have to select people at random and assign to then at random the treatment "smoke" or "don't smoke" and then measure the possible consequences. But preventing\r\n people from smoking is difficult and forcing them to smoke is unethical. Hence the importance of figuring out how far we can get in inferring causality without carrying out experiments.
\r\n4) Methods for inferring causal direction. Using observational data (of the kind we provide in the challenge) some of the state-of-the-art methods fail in cases that are even obvious to the human eye (for instance\r\n a non invertible function). We want to get at least all these obvious cases automatically and expect the challenge participants to exploit asymmetries in A->B and A
", "rawMarkdown": "", "votes": null }, { "id": "22011", "postDate": "04/02/2013 18:11:06", "content": "[quote=Ling Cheng;22007]
\r\nThanks Isabelle!
\r\nBtw, I don't think we have access to the google docs link you posted.
\r\n[/quote]
\r\nhttps://docs.google.com/a/chalearn.org/file/d/0B08QS7nJpK7mNUxudDdKY0MxRlE/edit
\r\nPlease try again, it should work now.
", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 21898, "author_name": "iguyon", "author_url": "", "post_date": "03/31/2013 14:41:41", "content": "No, this means smoking=A, lung cancer=B (only 2 features). You get data points in a 2-dimensional space {2092, 5651}, {1143, 4449}, {390, 4012}, etc. That is you can make a plot of B vs A. The question is: Is B a cause of A or A a cause of B.
\r\n- A is a cause of B means here that B=f(A, noise) [denoted A -> B]
\r\n- A is a cause of A means A=f(B, noise) [denoted A <- B]
\r\nBut a dependency between A and B can also be explained by a common cause Z:
\r\nA=f(Z, noise) and B=f(Z, noise) [denoted A - B]
\r\nAlso, some pairs of variables are intependent (denoted A | B) but because of the limited number of data points, they may coincidentally look dependent.
\r\nSee Slide presentation: overview of the challenge. There are a few examples there.
\r\nThere is a bibliography at http://www.causality.inf.ethz.ch/cause-effect.php?page=help.
", "votes": null, "replies": [] }, { "id": 21903, "author_name": "kimsimonsen", "author_url": "", "post_date": "03/31/2013 16:47:16", "content": "[quote=Isabelle;21898]
\r\nSee Slide presentation: overview of the challenge. There are a few examples there.
\r\n[/quote]
\r\nThank you the scatter plot illustrations really helps a lot in understanding the data.
", "votes": null, "replies": [] }, { "id": 21925, "author_name": "jc3626439", "author_url": "", "post_date": "04/01/2013 04:38:28", "content": "Isabelle, Is that a typo at the beginning of
\r\n- A is a cause of A means A=f(B, noise) [denoted A <- B]
\r\nin your response to LoveSaves?
", "votes": null, "replies": [] }, { "id": 21934, "author_name": "maddi24232", "author_url": "", "post_date": "04/01/2013 15:06:36", "content": "Yes, sorry, I meant B is a cause of A:
\r\n- B is a cause of A means A=f(B, noise) [denoted A <- B]
", "votes": null, "replies": [] }, { "id": 21969, "author_name": "iguyon", "author_url": "", "post_date": "04/02/2013 00:56:10", "content": "Your feed-back prompted ne to add a quiz on the first page, thanks.
\r\nI also provided some references at: http://www.causality.inf.ethz.ch/cause-effect.php?page=help
", "votes": null, "replies": [] }, { "id": 21981, "author_name": "lingcheng", "author_url": "", "post_date": "04/02/2013 07:17:39", "content": "Thank you for the Slide Presentation.
\r\nIt seems like there are a few mix-ups on the explanations on the right of the scatter plot examples as well?
\r\nE.g. Slide 12, A=Ages, B=Wages should be A->B not A<-B?
", "votes": null, "replies": [] }, { "id": 21989, "author_name": "domcastro", "author_url": "", "post_date": "04/02/2013 11:56:29", "content": "Bit confused as well. Doesn't Altitude determine temperature A <- B in the slide, rather than temperature determines altitude? Also, as with the commenter above, doesn't Age determine wages? so A -> B
\r\n\r\n
Both of these slides are opposite to what is expected?
", "votes": null, "replies": [] }, { "id": 21990, "author_name": "liubenyuan", "author_url": "", "post_date": "04/02/2013 12:21:34", "content": "even bit confused !!
\r\nI am using IGCI as the causal solver, as shown in the example of IGCI,
\r\nx = randn(100,1); y = sin(x); res = igci(x,y,2,1); disp(res); % with res>0
\r\nwhile, what about asin function? it can also be interpretate as x->y,
\r\nx = randn(100,1); y = asin(x); res = igci(x,y,2,1); disp(res); % with res<0
\r\nso what's the concrete definition of 'cause' ? can we assume that the inverse operator like\r\nasin is the inverse of causal?
", "votes": null, "replies": [] }, { "id": 22005, "author_name": "iguyon", "author_url": "", "post_date": "04/02/2013 17:13:48", "content": "[quote=Ling Cheng;21981]
\r\nThank you for the Slide Presentation.
\r\nIt seems like there are a few mix-ups on the explanations on the right of the scatter plot examples as well?
\r\nE.g. Slide 12, A=Ages, B=Wages should be A->B not A
\r\n[/quote]
\r\n\r\n
[quote=Domcastro;21989]
\r\nBit confused as well. Doesn't Altitude determine temperature AB
\r\n\r\n
Both of these slides are opposite to what is expected?
\r\n[/quote]
\r\nThank you for pointing this out, there is an error in the slides that I rectified:
\r\nB=altitude -> A=temperature
\r\nA=age -> B=wages
\r\nSorry about that.
\r\nhttps://docs.google.com/a/chalearn.org/file/d/0B08QS7nJpK7mNUxudDdKY0MxRlE/edit
", "votes": null, "replies": [] }, { "id": 22007, "author_name": "lingcheng", "author_url": "", "post_date": "04/02/2013 17:57:07", "content": "Thanks Isabelle!
\r\nBtw, I don't think we have access to the google docs link you posted.
", "votes": null, "replies": [] }, { "id": 22009, "author_name": "iguyon", "author_url": "", "post_date": "04/02/2013 18:08:12", "content": "[quote=liubenyuan;21990]
\r\neven bit confused !!
\r\nI am using IGCI as the causal solver, as shown in the example of IGCI,
\r\nx = randn(100,1); y = sin(x); res = igci(x,y,2,1); disp(res); % with res>0
\r\nwhile, what about asin function? it can also be interpretate as x->y,
\r\nx = randn(100,1); y = asin(x); res = igci(x,y,2,1); disp(res); % with res
\r\nso what's the concrete definition of 'cause' ? can we assume that the inverse operator like\r\nasin is the inverse of causal?
\r\n[/quote]
\r\nYou are touching at the heart of the problem. But there are several issues here:
\r\n1) Definition of causality. For artificial data, if B was generated as a function of A and some other "noise" variables, then A is a cause of B, by definition. For real data, we have to assume that there is\r\n an underlying data generating process that is equivalent to a mathematical function achieving the same thing. In the dataset we provided, we determined ground truth with the knowledge we have of the variables (if there is a statisticall dependency between\r\n temperature and altitude of cities, we will say that altitude causes temperature because of what we know of what these variables mean and how they have been measured).
\r\n2) Identifiability.In either case (artificial or natural), it does not mean that the causal direction is identifiable with observational data. For instance, if A and the noise are Gaussian distributed and B\r\n = w * A + noise, then it is impossible to determine whether A -> B or A
\r\n3) Experimentation: the gold standard. The gold standard for determining causal relationships is experimentations: kick A and see what happens to B. This is different from just observing values. For instance,\r\n if there is a device with a needle whose position correlates with temperature, if it is a themometer, forcing the needle to a given position will not change temparature; if it is a thermostat, it will. Our end goal is actually to find factors on which we can\r\n act to trigger desired effects. We do not expect that we will always be able to do so by just observing without experimenting. In non identifiable cases, we will have to resort to experimenting (e.g. setting A to given values and measure B while randomizing\r\n over other factors and vice versa). For instance, in the smoking and lung cancer example, we would have to select people at random and assign to then at random the treatment "smoke" or "don't smoke" and then measure the possible consequences. But preventing\r\n people from smoking is difficult and forcing them to smoke is unethical. Hence the importance of figuring out how far we can get in inferring causality without carrying out experiments.
\r\n4) Methods for inferring causal direction. Using observational data (of the kind we provide in the challenge) some of the state-of-the-art methods fail in cases that are even obvious to the human eye (for instance\r\n a non invertible function). We want to get at least all these obvious cases automatically and expect the challenge participants to exploit asymmetries in A->B and A
", "votes": null, "replies": [] }, { "id": 22011, "author_name": "iguyon", "author_url": "", "post_date": "04/02/2013 18:11:06", "content": "[quote=Ling Cheng;22007]
\r\nThanks Isabelle!
\r\nBtw, I don't think we have access to the google docs link you posted.
\r\n[/quote]
\r\nhttps://docs.google.com/a/chalearn.org/file/d/0B08QS7nJpK7mNUxudDdKY0MxRlE/edit
\r\nPlease try again, it should work now.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "21892": "", "21898": "", "21903": "", "21925": "", "21934": "", "21969": "", "21981": "", "21989": "", "21990": "", "22005": "", "22007": "", "22009": "", "22011": "" }, "source": "meta" }