Scientific competitions are very interesting as we have to learn a lot to be competitive when we are not scientists in the field of the competition. But at the same time, someone with deep knowledge in the field can leapfrog seasoned ML/DL practitioners.
\n\nHere is a short overview of several scientific competitions I entered
\n\nPlasticc, won by a scientist with shallow ML knowledge\nLanl, top of LB seems to be ML/DL specialists\nTrackml, won by scientist, 4 scientist teams in top 5
\n\nWill we see scientists domination as in Plasticc or Trackml, or not? Current top of LB includes both scientists and ML/DL specialists.
\n\nI'm not including scientific image competitions such as the two ongoing ones on Pneumothorax and diabetes, but a similar analysis would be welcome.
", "messages": [ { "id": 570651, "postDate": "2019-07-08T15:24:04.683Z", "content": "Scientific competitions are very interesting as we have to learn a lot to be competitive when we are not scientists in the field of the competition. But at the same time, someone with deep knowledge in the field can leapfrog seasoned ML/DL practitioners.
\n\nHere is a short overview of several scientific competitions I entered
\n\nPlasticc, won by a scientist with shallow ML knowledge\nLanl, top of LB seems to be ML/DL specialists\nTrackml, won by scientist, 4 scientist teams in top 5
\n\nWill we see scientists domination as in Plasticc or Trackml, or not? Current top of LB includes both scientists and ML/DL specialists.
\n\nI'm not including scientific image competitions such as the two ongoing ones on Pneumothorax and diabetes, but a similar analysis would be welcome.
", "rawMarkdown": "Scientific competitions are very interesting as we have to learn a lot to be competitive when we are not scientists in the field of the competition. But at the same time, someone with deep knowledge in the field can leapfrog seasoned ML/DL practitioners. \n\nHere is a short overview of several scientific competitions I entered \n\nPlasticc, won by a scientist with shallow ML knowledge\nLanl, top of LB seems to be ML/DL specialists\nTrackml, won by scientist, 4 scientist teams in top 5\n\nWill we see scientists domination as in Plasticc or Trackml, or not? Current top of LB includes both scientists and ML/DL specialists.\n\nI'm not including scientific image competitions such as the two ongoing ones on Pneumothorax and diabetes, but a similar analysis would be welcome.", "votes": 23 }, { "id": 573599, "postDate": "2019-07-12T13:25:55.257Z", "content": "Jaechang Lim, currently #1 LB: -2.686, is a molecular design expert. Such experts clearly have a very sizable head start in this contest.
\n\nFrom one of his papers: \"... Here we show that the performance of graph convolutional networks (GCNs) for the prediction of molecular properties can be improved by incorporating attention and gate mechanisms. The attention mechanism enables a GCN to identify atoms in different environments....\"
\n\nGithub Model used in the paper above
", "rawMarkdown": "Jaechang Lim, currently #1 LB: -2.686, is a molecular design expert. Such experts clearly have a very sizable head start in this contest.\n\nFrom one of his papers: \"... Here we show that the performance of **graph convolutional networks (GCNs)** for the prediction of molecular properties can be improved by incorporating attention and gate mechanisms. The attention mechanism enables a GCN to **identify atoms in different environments**....\"\n\n[Github Model used in the paper above](https://github.com/seongokryu/augmented-gcn)", "votes": 14 }, { "id": 571350, "postDate": "2019-07-09T14:15:17.883Z", "content": "The more interesting question, with 2 more month to go, whats the score limit :D
", "rawMarkdown": "The more interesting question, with 2 more month to go, whats the score limit :D", "votes": 7, "replies": [ { "id": 571369, "postDate": "2019-07-09T14:33:46.070Z", "content": "Maybe we could extrapolate LB scores dependence on time :D
", "rawMarkdown": "Maybe we could extrapolate [LB scores dependence on time](https://www.kaggle.com/robikscube/the-race-to-predict-molecular-properties#All-competitors-LB-Position-over-Time) :D", "votes": 1 }, { "id": 571379, "postDate": "2019-07-09T14:44:56.027Z", "content": "\n\n\nThe more interesting question
\n
Sure, you have both ML/DL experts and scientist on your team, my question is irrelevant to you ;)
", "rawMarkdown": "> The more interesting question\n\nSure, you have both ML/DL experts and scientist on your team, my question is irrelevant to you ;)" }, { "id": 571602, "postDate": "2019-07-09T19:55:20.643Z", "content": "\n\n\nML/DL experts and scientist
\n
Anyways boundaries are quite blurry. Sure we have a chemist, but also other team members (which you might call ML experts) have PhD in physics or mathematics and have some understanding of quantum mechanics. \nAfter all, it's not rocket science. ;)
", "rawMarkdown": "> ML/DL experts and scientist\n\nAnyways boundaries are quite blurry. Sure we have a chemist, but also other team members (which you might call ML experts) have PhD in physics or mathematics and have some understanding of quantum mechanics. \nAfter all, it's not rocket science. ;)\n\n", "votes": 2 }, { "id": 571605, "postDate": "2019-07-09T20:06:13.640Z", "content": "@christofhenkel you're right, rocket science is a good deal easier...
", "rawMarkdown": "@christofhenkel you're right, rocket science is a good deal easier...", "votes": 5 }, { "id": 571615, "postDate": "2019-07-09T20:36:41.073Z", "content": "ok, you have ML/DL experts and ScientistS on the team ;)
", "rawMarkdown": "ok, you have ML/DL experts and ScientistS on the team ;)" }, { "id": 572869, "postDate": "2019-07-11T14:06:26.613Z", "content": "May be -2.966 👀 . I will be trying my luck on lightGBM and Blends😋
", "rawMarkdown": "May be -2.966 👀 . I will be trying my luck on lightGBM and Blends😋 ", "votes": 1 }, { "id": 574111, "postDate": "2019-07-13T08:51:44.127Z", "content": "It seems that gold line will be definetely below -2.5. And the winner is probably going to be below -3.
", "rawMarkdown": "It seems that gold line will be definetely below -2.5. And the winner is probably going to be below -3.", "votes": 3 }, { "id": 590782, "postDate": "2019-08-02T15:27:49.007Z", "content": "Ok, now I update my guess. -2.7 will be the limit for gold, -3 will be the limit for money prize.
", "rawMarkdown": "Ok, now I update my guess. -2.7 will be the limit for gold, -3 will be the limit for money prize." }, { "id": 590865, "postDate": "2019-08-02T18:14:24.090Z", "content": "I think gold limit will be below -2.8
", "rawMarkdown": "I think gold limit will be below -2.8", "votes": 1 }, { "id": 602880, "postDate": "2019-08-19T15:55:34.997Z", "content": "\n\n\nOk, now I update my guess. -2.7 will be the limit for gold, -3 will be the limit for money prize
\n
seems like people are getting ambitious. looks like it will be hard to get gold with -3
", "rawMarkdown": "> Ok, now I update my guess. -2.7 will be the limit for gold, -3 will be the limit for money prize\n\nseems like people are getting ambitious. looks like it will be hard to get gold with -3 ", "votes": 2 } ] }, { "id": 570680, "postDate": "2019-07-08T16:12:26.413Z", "content": "In the Dark Worlds competition (gravitational lensing), the money went to 1) a phd in econometrics, 2) a ML researcher, and 3) (not sure the background).
", "rawMarkdown": "In the Dark Worlds competition (gravitational lensing), the money went to 1) a phd in econometrics, 2) a ML researcher, and 3) (not sure the background).", "votes": 6, "replies": [ { "id": 570732, "postDate": "2019-07-08T17:54:24.350Z", "content": "When was this one? There was also the Higgs Boson competition. Both happened before I became active I guess.
", "rawMarkdown": "When was this one? There was also the Higgs Boson competition. Both happened before I became active I guess.", "votes": 2 }, { "id": 570843, "postDate": "2019-07-08T20:33:46.790Z", "content": "The competition was in 2012. It was the 2nd competition I participated in, and was quite interesting. If I remember correctly, 2 of the 3 winners used Bayesian methods.
", "rawMarkdown": "The competition was in 2012. It was the 2nd competition I participated in, and was quite interesting. If I remember correctly, 2 of the 3 winners used Bayesian methods.", "votes": 2 }, { "id": 570946, "postDate": "2019-07-09T00:41:52.580Z", "content": "@inversion thanks for the information! I have to check those comp 😄 (I had no idea about those interesting competitions)
", "rawMarkdown": "@inversion thanks for the information! I have to check those comp 😄 (I had no idea about those interesting competitions)", "votes": 1 } ] }, { "id": 570678, "postDate": "2019-07-08T16:10:33.390Z", "content": "Maybe. But I am really looking forward for the solutions/ideas coming from team 4 GM and the brain. It is interesting to see what is their magic sauce. This is probably the strongest team I have seen this year in any competition. As an analogy imagine Messi, Neymar, Mbappe and Mo Salah playing on same side. Heng CherKeng is also taking outstanding approach will not be surprised if he wins this one.
\n\n\nThis is probably the strongest team I have seen this year in any competition.
\n
Interesting you say that for a team that has only one competition GM on it.
\n\nI am not downplaying that team by any mean, but why are you downplaying the solo teams that are achieving way better scores?
", "rawMarkdown": "> This is probably the strongest team I have seen this year in any competition.\n\nInteresting you say that for a team that has only one competition GM on it.\n\nI am not downplaying that team by any mean, but why are you downplaying the solo teams that are achieving way better scores? \n", "votes": 2 }, { "id": 570772, "postDate": "2019-07-08T18:47:01.947Z", "content": "With all regards to solo participants, after taking part as a solo team in some 20ish competitions myself. I feel working in a team is way better. The synergy adds and one learns a lot of things in a team from real world point of view we all have teams in our offices. I feel if one is solo in the last week they just feel like getting short of breath whereas if one is in a team usually they keep pushing to the last moment.
\n\nEspecially if one is in very far timezones lets say in a team of three if one is in North America, another is in Europe and last one in East Asia or Australia. Individual participant can't have 24*7 access to Kaggle (even though we sleep a little). But a team could have 100% uptime. This could be a deciding factor if it is kernels only competition trend which we are seeing a lot these days.
\n\nAssumption made: The skill level, knowledge and expertise of participants follow a Gaussian Curve. Interestingly I came across a not so known kaggler he has only competed as a solo participant and is a competitions Grandmaster, I am just imagining the quality of this kaggler would be so good, but unfortunately he looks like not to be so much active these days. #Respect.
", "rawMarkdown": "With all regards to solo participants, after taking part as a solo team in some 20ish competitions myself. I feel working in a team is way better. The synergy adds and one learns a lot of things in a team from real world point of view we all have teams in our offices. I feel if one is solo in the last week they just feel like getting short of breath whereas if one is in a team usually they keep pushing to the last moment.\n\nEspecially if one is in very far timezones lets say in a team of three if one is in North America, another is in Europe and last one in East Asia or Australia. Individual participant can't have 24*7 access to Kaggle (even though we sleep a little). But a team could have 100% uptime. This could be a deciding factor if it is kernels only competition trend which we are seeing a lot these days.\n\nAssumption made: The skill level, knowledge and expertise of participants follow a Gaussian Curve. Interestingly I came across a not so known kaggler he has only competed as a solo participant and is a competitions Grandmaster, I am just imagining the quality of this kaggler would be so good, but unfortunately he looks like not to be so much active these days. #Respect.", "votes": 1 }, { "id": 570947, "postDate": "2019-07-09T00:44:59.357Z", "content": "@jaechang is 5th, 7 attempts , alone ... I'm not sure that the more GM the stronger the team ;)
", "rawMarkdown": "@jaechang is 5th, 7 attempts , alone ... I'm not sure that the more GM the stronger the team ;)", "votes": 2 }, { "id": 570954, "postDate": "2019-07-09T01:08:08.817Z", "content": "Never know non-competition grandmasters can have 5 yellow stars in their profile pictures.
", "rawMarkdown": "Never know non-competition grandmasters can have 5 yellow stars in their profile pictures. " }, { "id": 571044, "postDate": "2019-07-09T05:23:08.530Z", "content": "5 stars are for grandmasters, not just competition grandmasters.
", "rawMarkdown": "5 stars are for grandmasters, not just competition grandmasters.", "votes": 1 }, { "id": 571185, "postDate": "2019-07-09T09:15:25.293Z", "content": "@cyberia Nice comparison :)
", "rawMarkdown": "@cyberia Nice comparison :)\n " }, { "id": 571224, "postDate": "2019-07-09T10:28:27.617Z", "content": "@cyberia teaming has its challenges too as you'll discover when you team. You can get removed because of a team mate not complying with competition rules. You can get passive team mates. You can get team mates that use all the team daily allocation for submissions. But you can also team with great people and achieve a better result than if you were solo, and learn a lot along the way. I experienced all these, the good and the bad, but overall teaming is positive.
", "rawMarkdown": "@cyberia teaming has its challenges too as you'll discover when you team. You can get removed because of a team mate not complying with competition rules. You can get passive team mates. You can get team mates that use all the team daily allocation for submissions. But you can also team with great people and achieve a better result than if you were solo, and learn a lot along the way. I experienced all these, the good and the bad, but overall teaming is positive. ", "votes": 1 } ] }, { "id": 573184, "postDate": "2019-07-11T22:40:16.663Z", "content": "From the point of view of a chemistry major who taught college level chemistry, I am fairly sure that my domain knowledge will be of limited help here. The reason for this is that NMR spectrum prediction is a very specialized field that is not taught in most undergraduate courses. What usually taught is simple heuristics of correlating neighborhood surrounding H or C with chemical shifts detected by NMR.
\n\nI would not totally discount the usefulness of chemistry knowledge for this problem. Surely, spending several years learning and thinking about molecules and functional group interactions should help generate interesting ideas. Having said this, a winning team without a chemist (or a related major) won't surprise me a bit.
", "rawMarkdown": "From the point of view of a chemistry major who taught college level chemistry, I am fairly sure that my domain knowledge will be of limited help here. The reason for this is that NMR spectrum prediction is a very specialized field that is not taught in most undergraduate courses. What usually taught is simple heuristics of correlating neighborhood surrounding H or C with chemical shifts detected by NMR.\n\nI would not totally discount the usefulness of chemistry knowledge for this problem. Surely, spending several years learning and thinking about molecules and functional group interactions should help generate interesting ideas. Having said this, a winning team without a chemist (or a related major) won't surprise me a bit.", "votes": 3 }, { "id": 589815, "postDate": "2019-08-01T12:16:04.850Z", "content": "With one month to go I checked a bit on the top 10 teams. Here are the ones I could identify as scientific experts in this field:
\n\n2 Jaechang Lim, see for instance https://arxiv.org/abs/1806.05805\n3 Tony Y. , see for instance https://tony-y.github.io/cgnn/\n5 Cédric Bény, see for instance http://qit.hanyang.ac.kr/member-cedric.html\n9 Boris D. see for instance https://www.linkedin.com/in/boris-dorado/\n10 Eric Renault, see for instance https://www.univ-nantes.fr/version-francaise/eric-renault-4073.kjsp
\n\nThere may be more as I could not identify some people.
", "rawMarkdown": "With one month to go I checked a bit on the top 10 teams. Here are the ones I could identify as scientific experts in this field:\n\n2 Jaechang Lim, see for instance https://arxiv.org/abs/1806.05805\n3 Tony Y. , see for instance https://tony-y.github.io/cgnn/\n5 Cédric Bény, see for instance http://qit.hanyang.ac.kr/member-cedric.html\n9 Boris D. see for instance https://www.linkedin.com/in/boris-dorado/\n10 Eric Renault, see for instance https://www.univ-nantes.fr/version-francaise/eric-renault-4073.kjsp\n\nThere may be more as I could not identify some people.\n\n", "votes": 4, "replies": [ { "id": 599660, "postDate": "2019-08-15T08:26:35.733Z", "content": "really impressive ... I can't wait to see how you are competing against that! :)
", "rawMarkdown": "really impressive ... I can't wait to see how you are competing against that! :)" }, { "id": 600567, "postDate": "2019-08-16T09:36:45.020Z", "content": "We try but it's hard ;)
", "rawMarkdown": "We try but it's hard ;)" } ] }, { "id": 609917, "postDate": "2019-08-28T08:27:13.410Z", "content": "I also suspect top 2 team to have specialists on board. The only one with identifiable name, Jonathan Mailoa, is certainly deep in chemistry: https://www.linkedin.com/in/jonathan-mailoa-b9572327
\n\nThis is to say that my prediction will probably be at least half true.
", "rawMarkdown": "I also suspect top 2 team to have specialists on board. The only one with identifiable name, Jonathan Mailoa, is certainly deep in chemistry: https://www.linkedin.com/in/jonathan-mailoa-b9572327\n\nThis is to say that my prediction will probably be at least half true.", "votes": 1 }, { "id": 574330, "postDate": "2019-07-13T16:26:50.283Z", "content": "There must be one more data science competition that who is going to win. As i guess that is going to be most interesting 😀 😂
", "rawMarkdown": "There must be one more data science competition that who is going to win. As i guess that is going to be most interesting 😀 😂 ", "votes": 1 }, { "id": 572871, "postDate": "2019-07-11T14:12:57.313Z", "content": "Oncle CPMP, are you jumping in ?
", "rawMarkdown": "Oncle CPMP, are you jumping in ?", "votes": 1, "replies": [ { "id": 572878, "postDate": "2019-07-11T14:21:34.207Z", "content": "I'm building models since start of the week, but I'm far from the published CV numbers... I'll submit this week end I think.
", "rawMarkdown": "I'm building models since start of the week, but I'm far from the published CV numbers... I'll submit this week end I think." }, { "id": 573037, "postDate": "2019-07-11T17:48:41.173Z", "content": "@cpmpml Let's team up! I boost you now, you boost me later 👍
\n\np.s. If I got lucky and you really up to do it, send me a message here or via LinkedIn.
", "rawMarkdown": "@cpmpml Let's team up! I boost you now, you boost me later 👍 \n\np.s. If I got lucky and you really up to do it, send me a message here or via LinkedIn.", "votes": 2 }, { "id": 573355, "postDate": "2019-07-12T06:53:47.020Z", "content": "Thanks, but I first want to see how far I can go alone. Give me a few weeks before I consider teaming possibly. I only started this Monday ;)
", "rawMarkdown": "Thanks, but I first want to see how far I can go alone. Give me a few weeks before I consider teaming possibly. I only started this Monday ;)", "votes": 2 }, { "id": 573707, "postDate": "2019-07-12T16:30:57.663Z", "content": "Well, it was worth a try ¯ \\ _ (ツ) _ / ¯
", "rawMarkdown": "Well, it was worth a try ¯ \\ _ (ツ) _ / ¯ ", "votes": 2 }, { "id": 581886, "postDate": "2019-07-22T13:52:56.323Z", "content": "I just made my first sub ;) Working 2 weeks without submitting is a new experience to me, and it is a bit stressing I must say. I wanted to be better than -2 and I barely made it :D
\n\nI admire those who only use a handful sub for a whole competition, not to mention @speculation single sub win in Expedia competition.
", "rawMarkdown": "I just made my first sub ;) Working 2 weeks without submitting is a new experience to me, and it is a bit stressing I must say. I wanted to be better than -2 and I barely made it :D\n\nI admire those who only use a handful sub for a whole competition, not to mention @speculation [single sub win in Expedia competition](https://www.kaggle.com/c/expedia-hotel-recommendations/leaderboard).", "votes": 1 }, { "id": 581928, "postDate": "2019-07-22T14:49:35.963Z", "content": "Incredible. For a newbie like me it is awesome.\nGood work!
", "rawMarkdown": "Incredible. For a newbie like me it is awesome.\nGood work!", "votes": 1 }, { "id": 581947, "postDate": "2019-07-22T15:14:20.777Z", "content": "Thanks!
", "rawMarkdown": "Thanks!", "votes": 1 }, { "id": 582192, "postDate": "2019-07-22T21:49:22.567Z", "content": "Winning with a single sub is my now confessed dream
", "rawMarkdown": "Winning with a single sub is my now confessed dream", "votes": 3 }, { "id": 582408, "postDate": "2019-07-23T06:35:32.587Z", "content": "Be first in a competition, whatever the number of subs, is my dream ;)
", "rawMarkdown": "Be first in a competition, whatever the number of subs, is my dream ;)", "votes": 1 } ] }, { "id": 609988, "postDate": "2019-08-28T09:52:12.300Z", "content": "You have missed a recent competition: Traveling Santa 2018 (optimization: traveling salesman problem (TSP)). The 1st team were a group of 2 celebrities: 2 professors dedicated their life just for TSP. Before they joined the competition (at around 1/3 timeline), everybody read their papers and used their baseline software (LHK, written in C). When they joined, they jumped straight to 1st position, and being there to the end, winning $17,000/$25,000 all prize money (including 1st prize, and 2 side prizes for the team occupying 1st position longest, and for the best kernel).
", "rawMarkdown": "You have missed a recent competition: Traveling Santa 2018 (optimization: traveling salesman problem (TSP)). The 1st team were a group of 2 celebrities: 2 professors dedicated their life just for TSP. Before they joined the competition (at around 1/3 timeline), everybody read their papers and used their baseline software (LHK, written in C). When they joined, they jumped straight to 1st position, and being there to the end, winning $17,000/$25,000 all prize money (including 1st prize, and 2 side prizes for the team occupying 1st position longest, and for the best kernel).", "votes": 2, "replies": [ { "id": 610021, "postDate": "2019-08-28T10:19:33.347Z", "content": "Right, but this was not to solve a scientific problem, it was a nice puzzle.
", "rawMarkdown": "Right, but this was not to solve a scientific problem, it was a nice puzzle.", "votes": 1 } ] }, { "id": 572566, "postDate": "2019-07-11T05:32:55.477Z", "content": "When I was reading and thinking about the components of J-coupling constant yesterday, it didn't take long to realize that any knowledge in nuclear chemistry will help me understand what aspects of a bond (the electrons, the orbitals, the atomic information?) contribute to the quantum computation of an e.g. Fermi contact interaction. While I certainly don't have adequate domain knowledge to construct to most probable features by hand, I am optimistic that high-dimensional methods and/or those with feature selection capability i.e. DL, trees might prove to find more patterns than the domain expert will assume. Or maybe I'm just blindly optimistic? I'm not super trying to win so -
", "rawMarkdown": "When I was reading and thinking about the components of J-coupling constant yesterday, it didn't take long to realize that any knowledge in nuclear chemistry will help me understand what aspects of a bond (the electrons, the orbitals, the atomic information?) contribute to the quantum computation of an e.g. Fermi contact interaction. While I certainly don't have adequate domain knowledge to construct to most probable features by hand, I am optimistic that high-dimensional methods and/or those with feature selection capability i.e. DL, trees might prove to find more patterns than the domain expert will assume. Or maybe I'm just blindly optimistic? I'm not super trying to win so -", "votes": 2 }, { "id": 573562, "postDate": "2019-07-12T12:22:35.330Z", "content": "I think you can not achieve a high score ( > - 2.00) without a high performance computer , especially with <= 16 Gb RAM. This is to me the common weak point of those Kaggle competitions, for the first weeks of the competition it is very interesting as you can invest various ideas, but then it boils down to putting as many features as you can put into your model. \nIn some competitions as this one, it can lead to a nonsense. The goal of this competition is to find the best alternative possible to heavy methods used by physicians to compute atom pair interactions. What is the point then to give a model that need more time to be 1) preprocessed (hundreds of features to be created) 2) runned ? .... In this competition there are no specific rules about that, but I think it should.
", "rawMarkdown": "I think you can not achieve a high score ( > - 2.00) without a high performance computer , especially with <= 16 Gb RAM. This is to me the common weak point of those Kaggle competitions, for the first weeks of the competition it is very interesting as you can invest various ideas, but then it boils down to putting as many features as you can put into your model. \nIn some competitions as this one, it can lead to a nonsense. The goal of this competition is to find the best alternative possible to heavy methods used by physicians to compute atom pair interactions. What is the point then to give a model that need more time to be 1) preprocessed (hundreds of features to be created) 2) runned ? .... In this competition there are no specific rules about that, but I think it should.", "replies": [ { "id": 574183, "postDate": "2019-07-13T10:33:37.493Z", "content": "For the time being I am using Kaggle kernels with score at -2.00 so I do hope it is possible to improve just using the kernels. Maybe this actually focuses you to explore better features or better representations of the atomic environment. In my opinion this is more useful than a larger computer, of course I could be wrong :)
", "rawMarkdown": "For the time being I am using Kaggle kernels with score at -2.00 so I do hope it is possible to improve just using the kernels. Maybe this actually focuses you to explore better features or better representations of the atomic environment. In my opinion this is more useful than a larger computer, of course I could be wrong :)", "votes": 7, "isDeleted": true }, { "id": 577648, "postDate": "2019-07-16T22:14:24.633Z", "content": "Would you be willing to collaborate? I have access to a NVIDIA DGX station, we can test the hypothesis :)
", "rawMarkdown": "Would you be willing to collaborate? I have access to a NVIDIA DGX station, we can test the hypothesis :) " }, { "id": 577966, "postDate": "2019-07-17T08:36:41.123Z", "content": "I think you're wrong. It may actually be the first competition where the number of features is so low :) If you read papers on graph neural nets on molecules, they usually have 10 features max.
", "rawMarkdown": "I think you're wrong. It may actually be the first competition where the number of features is so low :) If you read papers on graph neural nets on molecules, they usually have 10 features max.", "votes": 1 }, { "id": 577980, "postDate": "2019-07-17T08:46:18.383Z", "content": "\n\n\nIt may actually be the first competition where the number of features is so low :)
\n
You are almost right, but LANL competition had one feature ;)
\n\nLast Talking Data had only 5 features (and 350M rows).
", "rawMarkdown": "> It may actually be the first competition where the number of features is so low :)\n\nYou are almost right, but LANL competition had one feature ;) \n\nLast Talking Data had only 5 features (and 350M rows).", "votes": 1 }, { "id": 578003, "postDate": "2019-07-17T09:14:52.563Z", "content": "Yes I should have added that I'm new to Kaggle and I obviously know very few of the competitions that have taken place here :D
", "rawMarkdown": "Yes I should have added that I'm new to Kaggle and I obviously know very few of the competitions that have taken place here :D" }, { "id": 578076, "postDate": "2019-07-17T10:31:31.043Z", "content": "No problem, you're right in general: there are way more features. It seems that less features looks more challenging for many. Probably because it puts emphasis on feature engineering, or NN architecture engineering.
", "rawMarkdown": "No problem, you're right in general: there are way more features. It seems that less features looks more challenging for many. Probably because it puts emphasis on feature engineering, or NN architecture engineering." }, { "id": 578757, "postDate": "2019-07-18T05:37:49.637Z", "content": "Thanks for the offer 'navier_stokes. So far the Kaggle kernel is working fine and I am getting used to that setup.
", "rawMarkdown": "Thanks for the offer 'navier_stokes. So far the Kaggle kernel is working fine and I am getting used to that setup. ", "isDeleted": true } ] }, { "id": 570753, "postDate": "2019-07-08T18:30:15.513Z", "content": "I think domaine knowledge may have huge impact here.
", "rawMarkdown": "I think domaine knowledge may have huge impact here. " }, { "id": 597940, "postDate": "2019-08-12T23:55:39.413Z", "content": "~Half a month to go and TOP 10 and only Cédric Bény and Boris D. remain.
\n\nEDIT: Jaechang Lim just got back!
", "rawMarkdown": "~Half a month to go and TOP 10 and only Cédric Bény and Boris D. remain.\n\nEDIT: Jaechang Lim just got back! ", "votes": 1, "replies": [ { "id": 598145, "postDate": "2019-08-13T07:40:26.507Z", "content": "You forgot Jaechang Lim ... And Tony Y hasn't submitted for 15 days, he can create a surprise...
", "rawMarkdown": "You forgot Jaechang Lim ... And Tony Y hasn't submitted for 15 days, he can create a surprise...", "votes": 1 }, { "id": 599615, "postDate": "2019-08-15T07:25:21.760Z", "content": "Also, Youhan Lee and SungHwan Choi from current top 7 team are quantum chemists, see
\n\nhttps://www.researchgate.net/profile/Youhan_Lee. \nhttps://scholar.google.co.kr/citations?user=rskngtIAAAAJ&hl=en
", "rawMarkdown": "Also, Youhan Lee and SungHwan Choi from current top 7 team are quantum chemists, see \n\nhttps://www.researchgate.net/profile/Youhan_Lee. \nhttps://scholar.google.co.kr/citations?user=rskngtIAAAAJ&hl=en", "votes": 1 }, { "id": 599616, "postDate": "2019-08-15T07:25:52.450Z", "content": "The race is still open IMHO ;)
", "rawMarkdown": "The race is still open IMHO ;)", "votes": 2 }, { "id": 602881, "postDate": "2019-08-19T15:57:50.047Z", "content": "@cpmpml \nAs you found it, my background is from chemical engineering and computational chemistry.\nAnd, my ML background is from kaggle, especially, you ;) I've learned many things from you. \nVery fun competition ever!
", "rawMarkdown": "@cpmpml \nAs you found it, my background is from chemical engineering and computational chemistry.\nAnd, my ML background is from kaggle, especially, you ;) I've learned many things from you. \nVery fun competition ever!", "votes": 2 }, { "id": 602892, "postDate": "2019-08-19T16:04:17.940Z", "content": "> And, my ML background is from kaggle, especially, you ;) I've learned many things from you.
\n\nThanks! I hope the student will not best the master here ;)
", "rawMarkdown": "> And, my ML background is from kaggle, especially, you ;) I've learned many things from you. \n\nThanks! I hope the student will not best the master here ;)", "votes": 1 }, { "id": 602901, "postDate": "2019-08-19T16:12:25.063Z", "content": "Aren't we all your students, uncle?
", "rawMarkdown": "Aren't we all your students, uncle?", "votes": 1 }, { "id": 602914, "postDate": "2019-08-19T16:24:08.743Z", "content": "I'm a student of all here as well. I learned from Kaggle community as well most, if not all, what I know about ML practice .
", "rawMarkdown": "I'm a student of all here as well. I learned from Kaggle community as well most, if not all, what I know about ML practice .", "votes": 1 } ] }, { "id": 609916, "postDate": "2019-08-28T08:24:50.127Z", "content": "I just found that Andres Torrubia from top1 team has a member with knowledge of physics, having worked at CERN. Not true specialist of the domain, but some relevant background.
", "rawMarkdown": "I just found that Andres Torrubia from top1 team has a member with knowledge of physics, [having worked at CERN](https://www.youcheck.co/yc/Andres-Torrubia-553). Not true specialist of the domain, but some relevant background.\n\n", "replies": [ { "id": 610077, "postDate": "2019-08-28T11:48:41.817Z", "content": "CERN was my only job prior to becoming a serial entrepreneur, I am an electronics/telecom engineer... limited knowledge of particle physics :-)
", "rawMarkdown": "CERN was my only job prior to becoming a serial entrepreneur, I am an electronics/telecom engineer... limited knowledge of particle physics :-) ", "votes": 1 }, { "id": 610087, "postDate": "2019-08-28T12:03:03.807Z", "content": "Fair enough! Thanks for correcting me. And good luck for final result, you must be stressed to see your lead over the rest evaporate like this.
", "rawMarkdown": "Fair enough! Thanks for correcting me. And good luck for final result, you must be stressed to see your lead over the rest evaporate like this.", "votes": 1 } ] }, { "id": 603889, "postDate": "2019-08-20T19:17:51.207Z", "content": "Deep knowledge in the subject matter will be a deciding factor in this competition no doubt, but not absolute in this competition.
", "rawMarkdown": "Deep knowledge in the subject matter will be a deciding factor in this competition no doubt, but not absolute in this competition. ", "replies": [ { "id": 603911, "postDate": "2019-08-20T20:00:40.457Z", "content": "I hope deep knowledge in the subject matter does not matter as we have none on our team.
", "rawMarkdown": "I hope deep knowledge in the subject matter does not matter as we have none on our team.", "votes": 2 } ] }, { "id": 578740, "postDate": "2019-07-18T05:08:02.710Z", "content": "A candidate with domain knowledge and ML/DL expertise would win this competition.
", "rawMarkdown": "A candidate with domain knowledge and ML/DL expertise would win this competition." }, { "id": 570804, "postDate": "2019-07-08T19:27:16.747Z", "rawMarkdown": "", "isDeleted": true }, { "id": 570802, "postDate": "2019-07-08T19:26:40.057Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 573599, "author_name": "Mycroft Holmes", "author_url": "", "post_date": "2019-07-12T13:25:55.257000", "content": "Jaechang Lim, currently #1 LB: -2.686, is a molecular design expert. Such experts clearly have a very sizable head start in this contest.
\n\nFrom one of his papers: \"... Here we show that the performance of graph convolutional networks (GCNs) for the prediction of molecular properties can be improved by incorporating attention and gate mechanisms. The attention mechanism enables a GCN to identify atoms in different environments....\"
\n\nGithub Model used in the paper above
", "votes": 14, "replies": [] }, { "id": 571350, "author_name": "Dieter", "author_url": "", "post_date": "2019-07-09T14:15:17.883000", "content": "The more interesting question, with 2 more month to go, whats the score limit :D
", "votes": 7, "replies": [ { "id": 571369, "author_name": "somnambWl", "author_url": "", "post_date": "2019-07-09T14:33:46.070000", "content": "Maybe we could extrapolate LB scores dependence on time :D
", "votes": 1, "replies": [] }, { "id": 571379, "author_name": "CPMP", "author_url": "", "post_date": "2019-07-09T14:44:56.027000", "content": "\n\n\nThe more interesting question
\n
Sure, you have both ML/DL experts and scientist on your team, my question is irrelevant to you ;)
", "votes": 0, "replies": [] }, { "id": 571602, "author_name": "Dieter", "author_url": "", "post_date": "2019-07-09T19:55:20.643000", "content": "\n\n\nML/DL experts and scientist
\n
Anyways boundaries are quite blurry. Sure we have a chemist, but also other team members (which you might call ML experts) have PhD in physics or mathematics and have some understanding of quantum mechanics. \nAfter all, it's not rocket science. ;)
", "votes": 2, "replies": [] }, { "id": 571605, "author_name": "Stanislav Blinov", "author_url": "", "post_date": "2019-07-09T20:06:13.640000", "content": "@christofhenkel you're right, rocket science is a good deal easier...
", "votes": 5, "replies": [] }, { "id": 571615, "author_name": "CPMP", "author_url": "", "post_date": "2019-07-09T20:36:41.073000", "content": "ok, you have ML/DL experts and ScientistS on the team ;)
", "votes": 0, "replies": [] }, { "id": 572869, "author_name": "Genos", "author_url": "", "post_date": "2019-07-11T14:06:26.613000", "content": "May be -2.966 👀 . I will be trying my luck on lightGBM and Blends😋
", "votes": 1, "replies": [] }, { "id": 574111, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2019-07-13T08:51:44.127000", "content": "It seems that gold line will be definetely below -2.5. And the winner is probably going to be below -3.
", "votes": 3, "replies": [] }, { "id": 590782, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2019-08-02T15:27:49.007000", "content": "Ok, now I update my guess. -2.7 will be the limit for gold, -3 will be the limit for money prize.
", "votes": 0, "replies": [] }, { "id": 590865, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-02T18:14:24.090000", "content": "I think gold limit will be below -2.8
", "votes": 1, "replies": [] }, { "id": 602880, "author_name": "Dieter", "author_url": "", "post_date": "2019-08-19T15:55:34.997000", "content": "\n\n\nOk, now I update my guess. -2.7 will be the limit for gold, -3 will be the limit for money prize
\n
seems like people are getting ambitious. looks like it will be hard to get gold with -3
", "votes": 2, "replies": [] } ] }, { "id": 570680, "author_name": "inversion", "author_url": "", "post_date": "2019-07-08T16:12:26.413000", "content": "In the Dark Worlds competition (gravitational lensing), the money went to 1) a phd in econometrics, 2) a ML researcher, and 3) (not sure the background).
", "votes": 6, "replies": [ { "id": 570732, "author_name": "CPMP", "author_url": "", "post_date": "2019-07-08T17:54:24.350000", "content": "When was this one? There was also the Higgs Boson competition. Both happened before I became active I guess.
", "votes": 2, "replies": [] }, { "id": 570843, "author_name": "inversion", "author_url": "", "post_date": "2019-07-08T20:33:46.790000", "content": "The competition was in 2012. It was the 2nd competition I participated in, and was quite interesting. If I remember correctly, 2 of the 3 winners used Bayesian methods.
", "votes": 2, "replies": [] }, { "id": 570946, "author_name": "Nanashi", "author_url": "", "post_date": "2019-07-09T00:41:52.580000", "content": "@inversion thanks for the information! I have to check those comp 😄 (I had no idea about those interesting competitions)
", "votes": 1, "replies": [] } ] }, { "id": 570678, "author_name": "cyberia", "author_url": "", "post_date": "2019-07-08T16:10:33.390000", "content": "Maybe. But I am really looking forward for the solutions/ideas coming from team 4 GM and the brain. It is interesting to see what is their magic sauce. This is probably the strongest team I have seen this year in any competition. As an analogy imagine Messi, Neymar, Mbappe and Mo Salah playing on same side. Heng CherKeng is also taking outstanding approach will not be surprised if he wins this one.
\n\n\nThis is probably the strongest team I have seen this year in any competition.
\n
Interesting you say that for a team that has only one competition GM on it.
\n\nI am not downplaying that team by any mean, but why are you downplaying the solo teams that are achieving way better scores?
", "votes": 2, "replies": [] }, { "id": 570772, "author_name": "cyberia", "author_url": "", "post_date": "2019-07-08T18:47:01.947000", "content": "With all regards to solo participants, after taking part as a solo team in some 20ish competitions myself. I feel working in a team is way better. The synergy adds and one learns a lot of things in a team from real world point of view we all have teams in our offices. I feel if one is solo in the last week they just feel like getting short of breath whereas if one is in a team usually they keep pushing to the last moment.
\n\nEspecially if one is in very far timezones lets say in a team of three if one is in North America, another is in Europe and last one in East Asia or Australia. Individual participant can't have 24*7 access to Kaggle (even though we sleep a little). But a team could have 100% uptime. This could be a deciding factor if it is kernels only competition trend which we are seeing a lot these days.
\n\nAssumption made: The skill level, knowledge and expertise of participants follow a Gaussian Curve. Interestingly I came across a not so known kaggler he has only competed as a solo participant and is a competitions Grandmaster, I am just imagining the quality of this kaggler would be so good, but unfortunately he looks like not to be so much active these days. #Respect.
", "votes": 1, "replies": [] }, { "id": 570947, "author_name": "Nanashi", "author_url": "", "post_date": "2019-07-09T00:44:59.357000", "content": "@jaechang is 5th, 7 attempts , alone ... I'm not sure that the more GM the stronger the team ;)
", "votes": 2, "replies": [] }, { "id": 570954, "author_name": "Kha Vo", "author_url": "", "post_date": "2019-07-09T01:08:08.817000", "content": "Never know non-competition grandmasters can have 5 yellow stars in their profile pictures.
", "votes": 0, "replies": [] }, { "id": 571044, "author_name": "CPMP", "author_url": "", "post_date": "2019-07-09T05:23:08.530000", "content": "5 stars are for grandmasters, not just competition grandmasters.
", "votes": 1, "replies": [] }, { "id": 571185, "author_name": "Ordy ELGD", "author_url": "", "post_date": "2019-07-09T09:15:25.293000", "content": "@cyberia Nice comparison :)
", "votes": 0, "replies": [] }, { "id": 571224, "author_name": "CPMP", "author_url": "", "post_date": "2019-07-09T10:28:27.617000", "content": "@cyberia teaming has its challenges too as you'll discover when you team. You can get removed because of a team mate not complying with competition rules. You can get passive team mates. You can get team mates that use all the team daily allocation for submissions. But you can also team with great people and achieve a better result than if you were solo, and learn a lot along the way. I experienced all these, the good and the bad, but overall teaming is positive.
", "votes": 1, "replies": [] } ] }, { "id": 573184, "author_name": "Mycroft Holmes", "author_url": "", "post_date": "2019-07-11T22:40:16.663000", "content": "From the point of view of a chemistry major who taught college level chemistry, I am fairly sure that my domain knowledge will be of limited help here. The reason for this is that NMR spectrum prediction is a very specialized field that is not taught in most undergraduate courses. What usually taught is simple heuristics of correlating neighborhood surrounding H or C with chemical shifts detected by NMR.
\n\nI would not totally discount the usefulness of chemistry knowledge for this problem. Surely, spending several years learning and thinking about molecules and functional group interactions should help generate interesting ideas. Having said this, a winning team without a chemist (or a related major) won't surprise me a bit.
", "votes": 3, "replies": [] }, { "id": 589815, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-01T12:16:04.850000", "content": "With one month to go I checked a bit on the top 10 teams. Here are the ones I could identify as scientific experts in this field:
\n\n2 Jaechang Lim, see for instance https://arxiv.org/abs/1806.05805\n3 Tony Y. , see for instance https://tony-y.github.io/cgnn/\n5 Cédric Bény, see for instance http://qit.hanyang.ac.kr/member-cedric.html\n9 Boris D. see for instance https://www.linkedin.com/in/boris-dorado/\n10 Eric Renault, see for instance https://www.univ-nantes.fr/version-francaise/eric-renault-4073.kjsp
\n\nThere may be more as I could not identify some people.
", "votes": 4, "replies": [ { "id": 599660, "author_name": "Nanashi", "author_url": "", "post_date": "2019-08-15T08:26:35.733000", "content": "really impressive ... I can't wait to see how you are competing against that! :)
", "votes": 0, "replies": [] }, { "id": 600567, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-16T09:36:45.020000", "content": "We try but it's hard ;)
", "votes": 0, "replies": [] } ] }, { "id": 609917, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-28T08:27:13.410000", "content": "I also suspect top 2 team to have specialists on board. The only one with identifiable name, Jonathan Mailoa, is certainly deep in chemistry: https://www.linkedin.com/in/jonathan-mailoa-b9572327
\n\nThis is to say that my prediction will probably be at least half true.
", "votes": 1, "replies": [] }, { "id": 574330, "author_name": "Raju Kumar Mishra", "author_url": "", "post_date": "2019-07-13T16:26:50.283000", "content": "There must be one more data science competition that who is going to win. As i guess that is going to be most interesting 😀 😂
", "votes": 1, "replies": [] }, { "id": 572871, "author_name": "eagle4", "author_url": "", "post_date": "2019-07-11T14:12:57.313000", "content": "Oncle CPMP, are you jumping in ?
", "votes": 1, "replies": [ { "id": 572878, "author_name": "CPMP", "author_url": "", "post_date": "2019-07-11T14:21:34.207000", "content": "I'm building models since start of the week, but I'm far from the published CV numbers... I'll submit this week end I think.
", "votes": 0, "replies": [] }, { "id": 573037, "author_name": "Stanislav Blinov", "author_url": "", "post_date": "2019-07-11T17:48:41.173000", "content": "@cpmpml Let's team up! I boost you now, you boost me later 👍
\n\np.s. If I got lucky and you really up to do it, send me a message here or via LinkedIn.
", "votes": 2, "replies": [] }, { "id": 573355, "author_name": "CPMP", "author_url": "", "post_date": "2019-07-12T06:53:47.020000", "content": "Thanks, but I first want to see how far I can go alone. Give me a few weeks before I consider teaming possibly. I only started this Monday ;)
", "votes": 2, "replies": [] }, { "id": 573707, "author_name": "Stanislav Blinov", "author_url": "", "post_date": "2019-07-12T16:30:57.663000", "content": "Well, it was worth a try ¯ \\ _ (ツ) _ / ¯
", "votes": 2, "replies": [] }, { "id": 581886, "author_name": "CPMP", "author_url": "", "post_date": "2019-07-22T13:52:56.323000", "content": "I just made my first sub ;) Working 2 weeks without submitting is a new experience to me, and it is a bit stressing I must say. I wanted to be better than -2 and I barely made it :D
\n\nI admire those who only use a handful sub for a whole competition, not to mention @speculation single sub win in Expedia competition.
", "votes": 1, "replies": [] }, { "id": 581928, "author_name": "MDDIAZ", "author_url": "", "post_date": "2019-07-22T14:49:35.963000", "content": "Incredible. For a newbie like me it is awesome.\nGood work!
", "votes": 1, "replies": [] }, { "id": 581947, "author_name": "CPMP", "author_url": "", "post_date": "2019-07-22T15:14:20.777000", "content": "Thanks!
", "votes": 1, "replies": [] }, { "id": 582192, "author_name": "Andrés Miguel Torrubia Sáez", "author_url": "", "post_date": "2019-07-22T21:49:22.567000", "content": "Winning with a single sub is my now confessed dream
", "votes": 3, "replies": [] }, { "id": 582408, "author_name": "CPMP", "author_url": "", "post_date": "2019-07-23T06:35:32.587000", "content": "Be first in a competition, whatever the number of subs, is my dream ;)
", "votes": 1, "replies": [] } ] }, { "id": 609988, "author_name": "Kha Vo", "author_url": "", "post_date": "2019-08-28T09:52:12.300000", "content": "You have missed a recent competition: Traveling Santa 2018 (optimization: traveling salesman problem (TSP)). The 1st team were a group of 2 celebrities: 2 professors dedicated their life just for TSP. Before they joined the competition (at around 1/3 timeline), everybody read their papers and used their baseline software (LHK, written in C). When they joined, they jumped straight to 1st position, and being there to the end, winning $17,000/$25,000 all prize money (including 1st prize, and 2 side prizes for the team occupying 1st position longest, and for the best kernel).
", "votes": 2, "replies": [ { "id": 610021, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-28T10:19:33.347000", "content": "Right, but this was not to solve a scientific problem, it was a nice puzzle.
", "votes": 1, "replies": [] } ] }, { "id": 572566, "author_name": "Gunny Liu", "author_url": "", "post_date": "2019-07-11T05:32:55.477000", "content": "When I was reading and thinking about the components of J-coupling constant yesterday, it didn't take long to realize that any knowledge in nuclear chemistry will help me understand what aspects of a bond (the electrons, the orbitals, the atomic information?) contribute to the quantum computation of an e.g. Fermi contact interaction. While I certainly don't have adequate domain knowledge to construct to most probable features by hand, I am optimistic that high-dimensional methods and/or those with feature selection capability i.e. DL, trees might prove to find more patterns than the domain expert will assume. Or maybe I'm just blindly optimistic? I'm not super trying to win so -
", "votes": 2, "replies": [] }, { "id": 573562, "author_name": "MaxR2D", "author_url": "", "post_date": "2019-07-12T12:22:35.330000", "content": "I think you can not achieve a high score ( > - 2.00) without a high performance computer , especially with <= 16 Gb RAM. This is to me the common weak point of those Kaggle competitions, for the first weeks of the competition it is very interesting as you can invest various ideas, but then it boils down to putting as many features as you can put into your model. \nIn some competitions as this one, it can lead to a nonsense. The goal of this competition is to find the best alternative possible to heavy methods used by physicians to compute atom pair interactions. What is the point then to give a model that need more time to be 1) preprocessed (hundreds of features to be created) 2) runned ? .... In this competition there are no specific rules about that, but I think it should.
", "votes": 0, "replies": [ { "id": 574183, "author_name": "", "author_url": "", "post_date": "2019-07-13T10:33:37.493000", "content": "For the time being I am using Kaggle kernels with score at -2.00 so I do hope it is possible to improve just using the kernels. Maybe this actually focuses you to explore better features or better representations of the atomic environment. In my opinion this is more useful than a larger computer, of course I could be wrong :)
", "votes": 7, "replies": [] }, { "id": 577648, "author_name": "navier_stokes", "author_url": "", "post_date": "2019-07-16T22:14:24.633000", "content": "Would you be willing to collaborate? I have access to a NVIDIA DGX station, we can test the hypothesis :)
", "votes": 0, "replies": [] }, { "id": 577966, "author_name": "Boris D.", "author_url": "", "post_date": "2019-07-17T08:36:41.123000", "content": "I think you're wrong. It may actually be the first competition where the number of features is so low :) If you read papers on graph neural nets on molecules, they usually have 10 features max.
", "votes": 1, "replies": [] }, { "id": 577980, "author_name": "CPMP", "author_url": "", "post_date": "2019-07-17T08:46:18.383000", "content": "\n\n\nIt may actually be the first competition where the number of features is so low :)
\n
You are almost right, but LANL competition had one feature ;)
\n\nLast Talking Data had only 5 features (and 350M rows).
", "votes": 1, "replies": [] }, { "id": 578003, "author_name": "Boris D.", "author_url": "", "post_date": "2019-07-17T09:14:52.563000", "content": "Yes I should have added that I'm new to Kaggle and I obviously know very few of the competitions that have taken place here :D
", "votes": 0, "replies": [] }, { "id": 578076, "author_name": "CPMP", "author_url": "", "post_date": "2019-07-17T10:31:31.043000", "content": "No problem, you're right in general: there are way more features. It seems that less features looks more challenging for many. Probably because it puts emphasis on feature engineering, or NN architecture engineering.
", "votes": 0, "replies": [] }, { "id": 578757, "author_name": "", "author_url": "", "post_date": "2019-07-18T05:37:49.637000", "content": "Thanks for the offer 'navier_stokes. So far the Kaggle kernel is working fine and I am getting used to that setup.
", "votes": 0, "replies": [] } ] }, { "id": 570753, "author_name": "Serigne ", "author_url": "", "post_date": "2019-07-08T18:30:15.513000", "content": "I think domaine knowledge may have huge impact here.
", "votes": 0, "replies": [] }, { "id": 597940, "author_name": "Andrés Miguel Torrubia Sáez", "author_url": "", "post_date": "2019-08-12T23:55:39.413000", "content": "~Half a month to go and TOP 10 and only Cédric Bény and Boris D. remain.
\n\nEDIT: Jaechang Lim just got back!
", "votes": 1, "replies": [ { "id": 598145, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-13T07:40:26.507000", "content": "You forgot Jaechang Lim ... And Tony Y hasn't submitted for 15 days, he can create a surprise...
", "votes": 1, "replies": [] }, { "id": 599615, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-15T07:25:21.760000", "content": "Also, Youhan Lee and SungHwan Choi from current top 7 team are quantum chemists, see
\n\nhttps://www.researchgate.net/profile/Youhan_Lee. \nhttps://scholar.google.co.kr/citations?user=rskngtIAAAAJ&hl=en
", "votes": 1, "replies": [] }, { "id": 599616, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-15T07:25:52.450000", "content": "The race is still open IMHO ;)
", "votes": 2, "replies": [] }, { "id": 602881, "author_name": "Youhan Lee", "author_url": "", "post_date": "2019-08-19T15:57:50.047000", "content": "@cpmpml \nAs you found it, my background is from chemical engineering and computational chemistry.\nAnd, my ML background is from kaggle, especially, you ;) I've learned many things from you. \nVery fun competition ever!
", "votes": 2, "replies": [] }, { "id": 602892, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-19T16:04:17.940000", "content": "> And, my ML background is from kaggle, especially, you ;) I've learned many things from you.
\n\nThanks! I hope the student will not best the master here ;)
", "votes": 1, "replies": [] }, { "id": 602901, "author_name": "Dieter", "author_url": "", "post_date": "2019-08-19T16:12:25.063000", "content": "Aren't we all your students, uncle?
", "votes": 1, "replies": [] }, { "id": 602914, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-19T16:24:08.743000", "content": "I'm a student of all here as well. I learned from Kaggle community as well most, if not all, what I know about ML practice .
", "votes": 1, "replies": [] } ] }, { "id": 609916, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-28T08:24:50.127000", "content": "I just found that Andres Torrubia from top1 team has a member with knowledge of physics, having worked at CERN. Not true specialist of the domain, but some relevant background.
", "votes": 0, "replies": [ { "id": 610077, "author_name": "Andrés Miguel Torrubia Sáez", "author_url": "", "post_date": "2019-08-28T11:48:41.817000", "content": "CERN was my only job prior to becoming a serial entrepreneur, I am an electronics/telecom engineer... limited knowledge of particle physics :-)
", "votes": 1, "replies": [] }, { "id": 610087, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-28T12:03:03.807000", "content": "Fair enough! Thanks for correcting me. And good luck for final result, you must be stressed to see your lead over the rest evaporate like this.
", "votes": 1, "replies": [] } ] }, { "id": 603889, "author_name": "juanSolo", "author_url": "", "post_date": "2019-08-20T19:17:51.207000", "content": "Deep knowledge in the subject matter will be a deciding factor in this competition no doubt, but not absolute in this competition.
", "votes": 0, "replies": [ { "id": 603911, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-20T20:00:40.457000", "content": "I hope deep knowledge in the subject matter does not matter as we have none on our team.
", "votes": 2, "replies": [] } ] }, { "id": 578740, "author_name": "karthik bappudi", "author_url": "", "post_date": "2019-07-18T05:08:02.710000", "content": "A candidate with domain knowledge and ML/DL expertise would win this competition.
", "votes": 0, "replies": [] }, { "id": 570804, "author_name": "", "author_url": "", "post_date": "2019-07-08T19:27:16.747000", "content": "", "votes": 0, "replies": [] }, { "id": 570802, "author_name": "", "author_url": "", "post_date": "2019-07-08T19:26:40.057000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "570651": "Scientific competitions are very interesting as we have to learn a lot to be competitive when we are not scientists in the field of the competition. But at the same time, someone with deep knowledge in the field can leapfrog seasoned ML/DL practitioners. \n\nHere is a short overview of several scientific competitions I entered \n\nPlasticc, won by a scientist with shallow ML knowledge\nLanl, top of LB seems to be ML/DL specialists\nTrackml, won by scientist, 4 scientist teams in top 5\n\nWill we see scientists domination as in Plasticc or Trackml, or not? Current top of LB includes both scientists and ML/DL specialists.\n\nI'm not including scientific image competitions such as the two ongoing ones on Pneumothorax and diabetes, but a similar analysis would be welcome.", "573599": "Jaechang Lim, currently #1 LB: -2.686, is a molecular design expert. Such experts clearly have a very sizable head start in this contest.\n\nFrom one of his papers: \"... Here we show that the performance of **graph convolutional networks (GCNs)** for the prediction of molecular properties can be improved by incorporating attention and gate mechanisms. The attention mechanism enables a GCN to **identify atoms in different environments**....\"\n\n[Github Model used in the paper above](https://github.com/seongokryu/augmented-gcn)", "571350": "The more interesting question, with 2 more month to go, whats the score limit :D", "570680": "In the Dark Worlds competition (gravitational lensing), the money went to 1) a phd in econometrics, 2) a ML researcher, and 3) (not sure the background).", "570678": "Maybe. But I am really looking forward for the solutions/ideas coming from team `4 GM and the brain`. It is interesting to see what is their magic sauce. This is probably the strongest team I have seen this year in any competition. As an analogy imagine Messi, Neymar, Mbappe and Mo Salah playing on same side. Heng CherKeng is also taking outstanding approach will not be surprised if he wins this one.", "573184": "From the point of view of a chemistry major who taught college level chemistry, I am fairly sure that my domain knowledge will be of limited help here. The reason for this is that NMR spectrum prediction is a very specialized field that is not taught in most undergraduate courses. What usually taught is simple heuristics of correlating neighborhood surrounding H or C with chemical shifts detected by NMR.\n\nI would not totally discount the usefulness of chemistry knowledge for this problem. Surely, spending several years learning and thinking about molecules and functional group interactions should help generate interesting ideas. Having said this, a winning team without a chemist (or a related major) won't surprise me a bit.", "589815": "With one month to go I checked a bit on the top 10 teams. Here are the ones I could identify as scientific experts in this field:\n\n2 Jaechang Lim, see for instance https://arxiv.org/abs/1806.05805\n3 Tony Y. , see for instance https://tony-y.github.io/cgnn/\n5 Cédric Bény, see for instance http://qit.hanyang.ac.kr/member-cedric.html\n9 Boris D. see for instance https://www.linkedin.com/in/boris-dorado/\n10 Eric Renault, see for instance https://www.univ-nantes.fr/version-francaise/eric-renault-4073.kjsp\n\nThere may be more as I could not identify some people.\n\n", "609917": "I also suspect top 2 team to have specialists on board. The only one with identifiable name, Jonathan Mailoa, is certainly deep in chemistry: https://www.linkedin.com/in/jonathan-mailoa-b9572327\n\nThis is to say that my prediction will probably be at least half true.", "574330": "There must be one more data science competition that who is going to win. As i guess that is going to be most interesting 😀 😂 ", "572871": "Oncle CPMP, are you jumping in ?", "609988": "You have missed a recent competition: Traveling Santa 2018 (optimization: traveling salesman problem (TSP)). The 1st team were a group of 2 celebrities: 2 professors dedicated their life just for TSP. Before they joined the competition (at around 1/3 timeline), everybody read their papers and used their baseline software (LHK, written in C). When they joined, they jumped straight to 1st position, and being there to the end, winning $17,000/$25,000 all prize money (including 1st prize, and 2 side prizes for the team occupying 1st position longest, and for the best kernel).", "572566": "When I was reading and thinking about the components of J-coupling constant yesterday, it didn't take long to realize that any knowledge in nuclear chemistry will help me understand what aspects of a bond (the electrons, the orbitals, the atomic information?) contribute to the quantum computation of an e.g. Fermi contact interaction. While I certainly don't have adequate domain knowledge to construct to most probable features by hand, I am optimistic that high-dimensional methods and/or those with feature selection capability i.e. DL, trees might prove to find more patterns than the domain expert will assume. Or maybe I'm just blindly optimistic? I'm not super trying to win so -", "573562": "I think you can not achieve a high score ( > - 2.00) without a high performance computer , especially with <= 16 Gb RAM. This is to me the common weak point of those Kaggle competitions, for the first weeks of the competition it is very interesting as you can invest various ideas, but then it boils down to putting as many features as you can put into your model. \nIn some competitions as this one, it can lead to a nonsense. The goal of this competition is to find the best alternative possible to heavy methods used by physicians to compute atom pair interactions. What is the point then to give a model that need more time to be 1) preprocessed (hundreds of features to be created) 2) runned ? .... In this competition there are no specific rules about that, but I think it should.", "570753": "I think domaine knowledge may have huge impact here. ", "597940": "~Half a month to go and TOP 10 and only Cédric Bény and Boris D. remain.\n\nEDIT: Jaechang Lim just got back! ", "609916": "I just found that Andres Torrubia from top1 team has a member with knowledge of physics, [having worked at CERN](https://www.youcheck.co/yc/Andres-Torrubia-553). Not true specialist of the domain, but some relevant background.\n\n", "603889": "Deep knowledge in the subject matter will be a deciding factor in this competition no doubt, but not absolute in this competition. ", "578740": "A candidate with domain knowledge and ML/DL expertise would win this competition.", "570804": "", "570802": "" } }