Just realized another problem that most of data scientists don't know the high energy physics domain knowledge, which might have prevented good competitors from participating in this game. Understanding the kinematics is very important for feature engineering, and PDG is always the best reference http://pdg.lbl.gov/2014/reviews/rpp2014-rev-kinematics.pdf for example, we can check 46.4.3. for three-boday decay (tau decays to 3 muons where m1, m2, m3 are all muon mass) This reference only needs high school level physics and math, no need to worry too much about special relativity, so please don't be afraid of reading it.
", "messages": [ { "id": "93921", "postDate": "10/02/2015 19:22:59", "content": "Just realized another problem that most of data scientists don't know the high energy physics domain knowledge, which might have prevented good competitors from participating in this game. Understanding the kinematics is very important for feature engineering, and PDG is always the best reference http://pdg.lbl.gov/2014/reviews/rpp2014-rev-kinematics.pdf for example, we can check 46.4.3. for three-boday decay (tau decays to 3 muons where m1, m2, m3 are all muon mass) This reference only needs high school level physics and math, no need to worry too much about special relativity, so please don't be afraid of reading it.
", "rawMarkdown": "Just realized another problem that most of data scientists don't know the high energy physics domain knowledge, which might have prevented good competitors from participating in this game. Understanding the kinematics is very important for feature engineering, and PDG is always the best reference http://pdg.lbl.gov/2014/reviews/rpp2014-rev-kinematics.pdf for example, we can check 46.4.3. for three-boday decay (tau decays to 3 muons where m1, m2, m3 are all muon mass) This reference only needs high school level physics and math, no need to worry too much about special relativity, so please don't be afraid of reading it.", "votes": null }, { "id": "94986", "postDate": "10/03/2015 23:51:52", "content": "Hi phunter,
\n\nI've been exploring the possibility of engineering features based on high school physics/math. I wanted to ask - Based on the features provided, is there any way to obtain the four-momentum of each particle? I know we have the transverse momentum and the magnitude of the cartesian momentum, from which we can calculate p_z, but is there any way of calculating p_x or p_y, or the energy? Thanks for your help.
", "rawMarkdown": "Hi phunter,\r\n\r\nI've been exploring the possibility of engineering features based on high school physics/math. I wanted to ask - Based on the features provided, is there any way to obtain the four-momentum of each particle? I know we have the transverse momentum and the magnitude of the cartesian momentum, from which we can calculate p_z, but is there any way of calculating p_x or p_y, or the energy? Thanks for your help.", "votes": null }, { "id": "94991", "postDate": "10/04/2015 01:28:45", "content": "without the phi angle, it is hard to calculate p_x p_y.
", "rawMarkdown": "without the phi angle, it is hard to calculate p_x p_y.", "votes": null }, { "id": "94993", "postDate": "10/04/2015 01:40:19", "content": "[quote=phunter;94991]
\n\nwithout the phi angle, it is hard to calculate p_x p_y.
\n\n[/quote]
\n\nCan the phi information of every event be recorded by LHCb? I guess they just don't want to provide this information in csv files.
", "rawMarkdown": "[quote=phunter;94991]\r\n\r\nwithout the phi angle, it is hard to calculate p_x p_y.\r\n\r\n[/quote]\r\n\r\nCan the phi information of every event be recorded by LHCb? I guess they just don't want to provide this information in csv files.", "votes": null }, { "id": "94995", "postDate": "10/04/2015 02:54:28", "content": "[quote=Liucheng Wang;94993]
\n\n[quote=phunter;94991]
\n\nwithout the phi angle, it is hard to calculate p_x p_y.
\n\n[/quote]
\n\nCan the phi information of every event be recorded by LHCb? I guess they just don't want to provide this information in csv files.
\n\n[/quote]\nI guess that, If we have phi info, the invariant mass could be calculated (equation in PDG), plus the information that Ds (strange meson) mass is 1.97 GeV and tau mass is 1.77 GeV, one can easily tell if a sample is control or signal.
", "rawMarkdown": "[quote=Liucheng Wang;94993]\r\n\r\n[quote=phunter;94991]\r\n\r\nwithout the phi angle, it is hard to calculate p_x p_y.\r\n\r\n[/quote]\r\n\r\nCan the phi information of every event be recorded by LHCb? I guess they just don't want to provide this information in csv files.\r\n\r\n[/quote]\r\nI guess that, If we have phi info, the invariant mass could be calculated (equation in PDG), plus the information that Ds (strange meson) mass is 1.97 GeV and tau mass is 1.77 GeV, one can easily tell if a sample is control or signal.", "votes": null }, { "id": "95065", "postDate": "10/04/2015 18:59:27", "content": "[quote=phunter;94995]
\n\nI guess that, If we have phi info, the invariant mass could be calculated (equation in PDG), plus the information that Ds (strange meson) mass is 1.97 GeV and tau mass is 1.77 GeV, one can easily tell if a sample is control or signal.
\n\n[/quote]
\n\nIn fact, you can calculate the invariant mass of the three decay particles from the variables given in the dataset. From what I see, for 95% of the training set, this agrees within +/-1% with the mass variable (assuming the three decay products all are muons).
\n\nIt's not clear how this would help building a classifier however. You can not just use the mass as a discriminant variable for example, the Cramer-von Mises test ensures the classifier output does not strongly depend on the mass of the three particle system.
\n\nIronically, due to the lack of the phi angles for the three decay products, we can't calculate the invariant masses of the pairs m01, m12 and m02 if I'm not mistaken. Which is a pity because the competition documentation says that the Ds_s -> eta mu- nu_mu background can be suppressed by requiring the mass of any opposite sign muon pair to be greater than 450 MeV/c^2. Neither do we have the sign of the decay product charges in the files.
", "rawMarkdown": "[quote=phunter;94995]\r\n\r\nI guess that, If we have phi info, the invariant mass could be calculated (equation in PDG), plus the information that Ds (strange meson) mass is 1.97 GeV and tau mass is 1.77 GeV, one can easily tell if a sample is control or signal.\r\n\r\n[/quote]\r\n\r\nIn fact, you can calculate the invariant mass of the three decay particles from the variables given in the dataset. From what I see, for 95% of the training set, this agrees within +/-1% with the mass variable (assuming the three decay products all are muons).\r\n\r\nIt's not clear how this would help building a classifier however. You can not just use the mass as a discriminant variable for example, the Cramer-von Mises test ensures the classifier output does not strongly depend on the mass of the three particle system.\r\n\r\nIronically, due to the lack of the phi angles for the three decay products, we can't calculate the invariant masses of the pairs m01, m12 and m02 if I'm not mistaken. Which is a pity because the competition documentation says that the Ds_s -> eta mu- nu_mu background can be suppressed by requiring the mass of any opposite sign muon pair to be greater than 450 MeV/c^2. Neither do we have the sign of the decay product charges in the files.", "votes": null }, { "id": "95085", "postDate": "10/04/2015 22:21:26", "content": "@Andre Holzner do we have variables (e.g. M_3body) that in the dataset for calculating 3 body invariant mass, or some other variable for 3 body invariant mass which I ignored? If yes, with +/- 1% resolution, people can easily distinguish Ds and tau which can separate control samples and training samples, so passing Ks test becomes trivial.
\n\nAnd, yes, without phi info, there was no way to calculate m01, m12 or m02, but with other variables, xgboost and other strong classifier can have good separation from tau->3 mu to Ds_s->eta mu nu_mu. And I agree, it is hard without charge/sign of muons.
", "rawMarkdown": "Andre Holzner do we have variables (e.g. M_3body) that in the dataset for calculating 3 body invariant mass, or some other variable for 3 body invariant mass which I ignored? If yes, with +/- 1% resolution, people can easily distinguish Ds and tau which can separate control samples and training samples, so passing Ks test becomes trivial.\r\n\r\nAnd, yes, without phi info, there was no way to calculate m01, m12 or m02, but with other variables, xgboost and other strong classifier can have good separation from tau->3 mu to Ds_s->eta mu nu_mu. And I agree, it is hard without charge/sign of muons.", "votes": null } ], "comments": [ { "id": 94986, "author_name": "arjunsubway", "author_url": "", "post_date": "10/03/2015 23:51:52", "content": "Hi phunter,
\n\nI've been exploring the possibility of engineering features based on high school physics/math. I wanted to ask - Based on the features provided, is there any way to obtain the four-momentum of each particle? I know we have the transverse momentum and the magnitude of the cartesian momentum, from which we can calculate p_z, but is there any way of calculating p_x or p_y, or the energy? Thanks for your help.
", "votes": null, "replies": [] }, { "id": 94991, "author_name": "phunter", "author_url": "", "post_date": "10/04/2015 01:28:45", "content": "without the phi angle, it is hard to calculate p_x p_y.
", "votes": null, "replies": [] }, { "id": 94993, "author_name": "liuchengwang", "author_url": "", "post_date": "10/04/2015 01:40:19", "content": "[quote=phunter;94991]
\n\nwithout the phi angle, it is hard to calculate p_x p_y.
\n\n[/quote]
\n\nCan the phi information of every event be recorded by LHCb? I guess they just don't want to provide this information in csv files.
", "votes": null, "replies": [] }, { "id": 94995, "author_name": "phunter", "author_url": "", "post_date": "10/04/2015 02:54:28", "content": "[quote=Liucheng Wang;94993]
\n\n[quote=phunter;94991]
\n\nwithout the phi angle, it is hard to calculate p_x p_y.
\n\n[/quote]
\n\nCan the phi information of every event be recorded by LHCb? I guess they just don't want to provide this information in csv files.
\n\n[/quote]\nI guess that, If we have phi info, the invariant mass could be calculated (equation in PDG), plus the information that Ds (strange meson) mass is 1.97 GeV and tau mass is 1.77 GeV, one can easily tell if a sample is control or signal.
", "votes": null, "replies": [] }, { "id": 95065, "author_name": "holzner", "author_url": "", "post_date": "10/04/2015 18:59:27", "content": "[quote=phunter;94995]
\n\nI guess that, If we have phi info, the invariant mass could be calculated (equation in PDG), plus the information that Ds (strange meson) mass is 1.97 GeV and tau mass is 1.77 GeV, one can easily tell if a sample is control or signal.
\n\n[/quote]
\n\nIn fact, you can calculate the invariant mass of the three decay particles from the variables given in the dataset. From what I see, for 95% of the training set, this agrees within +/-1% with the mass variable (assuming the three decay products all are muons).
\n\nIt's not clear how this would help building a classifier however. You can not just use the mass as a discriminant variable for example, the Cramer-von Mises test ensures the classifier output does not strongly depend on the mass of the three particle system.
\n\nIronically, due to the lack of the phi angles for the three decay products, we can't calculate the invariant masses of the pairs m01, m12 and m02 if I'm not mistaken. Which is a pity because the competition documentation says that the Ds_s -> eta mu- nu_mu background can be suppressed by requiring the mass of any opposite sign muon pair to be greater than 450 MeV/c^2. Neither do we have the sign of the decay product charges in the files.
", "votes": null, "replies": [] }, { "id": 95085, "author_name": "phunter", "author_url": "", "post_date": "10/04/2015 22:21:26", "content": "@Andre Holzner do we have variables (e.g. M_3body) that in the dataset for calculating 3 body invariant mass, or some other variable for 3 body invariant mass which I ignored? If yes, with +/- 1% resolution, people can easily distinguish Ds and tau which can separate control samples and training samples, so passing Ks test becomes trivial.
\n\nAnd, yes, without phi info, there was no way to calculate m01, m12 or m02, but with other variables, xgboost and other strong classifier can have good separation from tau->3 mu to Ds_s->eta mu nu_mu. And I agree, it is hard without charge/sign of muons.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "93921": "Just realized another problem that most of data scientists don't know the high energy physics domain knowledge, which might have prevented good competitors from participating in this game. Understanding the kinematics is very important for feature engineering, and PDG is always the best reference http://pdg.lbl.gov/2014/reviews/rpp2014-rev-kinematics.pdf for example, we can check 46.4.3. for three-boday decay (tau decays to 3 muons where m1, m2, m3 are all muon mass) This reference only needs high school level physics and math, no need to worry too much about special relativity, so please don't be afraid of reading it.", "94986": "Hi phunter,\r\n\r\nI've been exploring the possibility of engineering features based on high school physics/math. I wanted to ask - Based on the features provided, is there any way to obtain the four-momentum of each particle? I know we have the transverse momentum and the magnitude of the cartesian momentum, from which we can calculate p_z, but is there any way of calculating p_x or p_y, or the energy? Thanks for your help.", "94991": "without the phi angle, it is hard to calculate p_x p_y.", "94993": "[quote=phunter;94991]\r\n\r\nwithout the phi angle, it is hard to calculate p_x p_y.\r\n\r\n[/quote]\r\n\r\nCan the phi information of every event be recorded by LHCb? I guess they just don't want to provide this information in csv files.", "94995": "[quote=Liucheng Wang;94993]\r\n\r\n[quote=phunter;94991]\r\n\r\nwithout the phi angle, it is hard to calculate p_x p_y.\r\n\r\n[/quote]\r\n\r\nCan the phi information of every event be recorded by LHCb? I guess they just don't want to provide this information in csv files.\r\n\r\n[/quote]\r\nI guess that, If we have phi info, the invariant mass could be calculated (equation in PDG), plus the information that Ds (strange meson) mass is 1.97 GeV and tau mass is 1.77 GeV, one can easily tell if a sample is control or signal.", "95065": "[quote=phunter;94995]\r\n\r\nI guess that, If we have phi info, the invariant mass could be calculated (equation in PDG), plus the information that Ds (strange meson) mass is 1.97 GeV and tau mass is 1.77 GeV, one can easily tell if a sample is control or signal.\r\n\r\n[/quote]\r\n\r\nIn fact, you can calculate the invariant mass of the three decay particles from the variables given in the dataset. From what I see, for 95% of the training set, this agrees within +/-1% with the mass variable (assuming the three decay products all are muons).\r\n\r\nIt's not clear how this would help building a classifier however. You can not just use the mass as a discriminant variable for example, the Cramer-von Mises test ensures the classifier output does not strongly depend on the mass of the three particle system.\r\n\r\nIronically, due to the lack of the phi angles for the three decay products, we can't calculate the invariant masses of the pairs m01, m12 and m02 if I'm not mistaken. Which is a pity because the competition documentation says that the Ds_s -> eta mu- nu_mu background can be suppressed by requiring the mass of any opposite sign muon pair to be greater than 450 MeV/c^2. Neither do we have the sign of the decay product charges in the files.", "95085": "Andre Holzner do we have variables (e.g. M_3body) that in the dataset for calculating 3 body invariant mass, or some other variable for 3 body invariant mass which I ignored? If yes, with +/- 1% resolution, people can easily distinguish Ds and tau which can separate control samples and training samples, so passing Ks test becomes trivial.\r\n\r\nAnd, yes, without phi info, there was no way to calculate m01, m12 or m02, but with other variables, xgboost and other strong classifier can have good separation from tau->3 mu to Ds_s->eta mu nu_mu. And I agree, it is hard without charge/sign of muons." }, "source": "meta" }