My score is not (yet) brilliant and I should explain my work so far. I've been too busy to give this competition my full attention, but over the summer my workload has decreased and I now have around 8 hours a day to focus on it - essentially a full-time job.
\n\nMy model currently uses only structural and mulliken charge features - none of Giba's features regarding Yukawa or Coulomb fields, and none of the other train datasets or QM9. I have engineered my own features to be invariant of the various symmetries present in molecules - most people seem to be using features that are dependant on atom_index numbers and x, y & z coordinates which I think are prone to identifying misleading correlations. I have reason to believe my features are strong, and my low score may be due to train/test feature distributions being skewed. I'm reducing their number accordingly. With 2JHH for example I can get a CV score of -2 using only 29 features.
I have used my knowledge of organic chemistry to engineer features that are consistent with factors that affect their molecular properties, including but not limited to, ring strain, dihedral angles, electronegative substituents and pi/sigma characteristics of organic bond types. I have done quite a bit of research on my own to to this end, and certainly have amassed some reading material that would be of interest to anyone! I have tried to make features that I didn't think would occur to most people, and I have been fortunate that many of these have proven to be important features for my models. I think I have a lot to offer any team.
\n\nAs for myself, I have a BSc in Physics and an MSc in an unrelated computing subject. I have worked previously with NMR data and while we are not allowed to use QM calculations for this competition my understanding may help. I have been studying ML for around 18 months and have broad experience with NNs and GBDT models. On that note, I feel that graphical neural networks are the most appropriate way forward with molecular data. The features I have engineered lend themselves well to this framework since they are rotationally invariant.
\n\nI am hardworking, competitive and driven by curiosity - for what it's worth, I've been told that I'm a pleasure to work with! If any team is interested, please message me over this weekend. I look forward to contributing.
", "messages": [ { "id": "585610", "postDate": "07/27/2019 18:18:39", "content": "My score is not (yet) brilliant and I should explain my work so far. I've been too busy to give this competition my full attention, but over the summer my workload has decreased and I now have around 8 hours a day to focus on it - essentially a full-time job.
\n\nMy model currently uses only structural and mulliken charge features - none of Giba's features regarding Yukawa or Coulomb fields, and none of the other train datasets or QM9. I have engineered my own features to be invariant of the various symmetries present in molecules - most people seem to be using features that are dependant on atom_index numbers and x, y & z coordinates which I think are prone to identifying misleading correlations. I have reason to believe my features are strong, and my low score may be due to train/test feature distributions being skewed. I'm reducing their number accordingly. With 2JHH for example I can get a CV score of -2 using only 29 features.
I have used my knowledge of organic chemistry to engineer features that are consistent with factors that affect their molecular properties, including but not limited to, ring strain, dihedral angles, electronegative substituents and pi/sigma characteristics of organic bond types. I have done quite a bit of research on my own to to this end, and certainly have amassed some reading material that would be of interest to anyone! I have tried to make features that I didn't think would occur to most people, and I have been fortunate that many of these have proven to be important features for my models. I think I have a lot to offer any team.
\n\nAs for myself, I have a BSc in Physics and an MSc in an unrelated computing subject. I have worked previously with NMR data and while we are not allowed to use QM calculations for this competition my understanding may help. I have been studying ML for around 18 months and have broad experience with NNs and GBDT models. On that note, I feel that graphical neural networks are the most appropriate way forward with molecular data. The features I have engineered lend themselves well to this framework since they are rotationally invariant.
\n\nI am hardworking, competitive and driven by curiosity - for what it's worth, I've been told that I'm a pleasure to work with! If any team is interested, please message me over this weekend. I look forward to contributing.
", "rawMarkdown": "My score is not (yet) brilliant and I should explain my work so far. I've been too busy to give this competition my full attention, but over the summer my workload has decreased and I now have around 8 hours a day to focus on it - essentially a full-time job. \n\nMy model currently uses only structural and mulliken charge features - none of Giba's features regarding Yukawa or Coulomb fields, and none of the other train datasets or QM9. I have engineered my own features to be invariant of the various symmetries present in molecules - most people seem to be using features that are dependant on `atom_index` numbers and x, y & z coordinates which I think are prone to identifying misleading correlations. I have reason to believe my features are strong, and my low score may be due to train/test feature distributions being skewed. I'm reducing their number accordingly. With 2JHH for example I can get a CV score of -2 using only 29 features.\n\nI have used my knowledge of organic chemistry to engineer features that are consistent with factors that affect their molecular properties, including but not limited to, ring strain, dihedral angles, electronegative substituents and pi/sigma characteristics of organic bond types. I have done quite a bit of research on my own to to this end, and certainly have amassed some reading material that would be of interest to anyone! I have tried to make features that I didn't think would occur to most people, and I have been fortunate that many of these have proven to be important features for my models. I think I have a lot to offer any team.\n\nAs for myself, I have a BSc in Physics and an MSc in an unrelated computing subject. I have worked previously with NMR data and while we are not allowed to use QM calculations for this competition my understanding may help. I have been studying ML for around 18 months and have broad experience with NNs and GBDT models. On that note, I feel that graphical neural networks are the most appropriate way forward with molecular data. The features I have engineered lend themselves well to this framework since they are rotationally invariant.\n\nI am hardworking, competitive and driven by curiosity - for what it's worth, I've been told that I'm a pleasure to work with! If any team is interested, please message me over this weekend. I look forward to contributing.", "votes": null }, { "id": "585636", "postDate": "07/27/2019 19:16:05", "content": "Very interested to work with you. currently have 3jhh at 0.12 MAE and 1jhc at 0.487 MAE. Training 3jhn and I at 0.0904 MAE right now. I havent tried to submit on LB for awhile though, so my current ranking does not reflect the updated CV's.
\n\ni private messaged you if you are interested
", "rawMarkdown": "Very interested to work with you. currently have 3jhh at 0.12 MAE and 1jhc at 0.487 MAE. Training 3jhn and I at 0.0904 MAE right now. I havent tried to submit on LB for awhile though, so my current ranking does not reflect the updated CV's.\n\ni private messaged you if you are interested", "votes": null }, { "id": "585680", "postDate": "07/27/2019 20:56:43", "content": "Also interested.\nMy model is lgbm, but I didn't work at this competition at all for 3 weeks already (I'm on vacation with no access to my computer). My CV (it's a bit old, but there were no significant changes):\ntype | log | mae |\n1JHC | -0.825 | 0.573 |\n1JHN | -1.650 | 0.296 |\n2JHC | -1.273 | 0.308 |\n2JHH | -2.138 | 0.151 |\n2JHN | -1.992 | 0.193 |\n3JHC | -1.841 | 0.433 |\n3JHH | -1.816 | 0.212 |\n3JHN | -2.32 | 0.157 |\nI'll be able to start participating again in 3-4 days (~1 Aug.).\nMy plan is to create features anew because for a moment I have a lot of trash in it and then to start building a NN.\nPlease let me know if you're interested.
", "rawMarkdown": "Also interested.\nMy model is lgbm, but I didn't work at this competition at all for 3 weeks already (I'm on vacation with no access to my computer). My CV (it's a bit old, but there were no significant changes):\ntype | log | mae |\n1JHC | -0.825 | 0.573 |\n1JHN | -1.650 | 0.296 |\n2JHC | -1.273 | 0.308 |\n2JHH | -2.138 | 0.151 |\n2JHN | -1.992 | 0.193 |\n3JHC | -1.841 | 0.433 |\n3JHH | -1.816 | 0.212 |\n3JHN | -2.32 | 0.157 |\nI'll be able to start participating again in 3-4 days (~1 Aug.).\nMy plan is to create features anew because for a moment I have a lot of trash in it and then to start building a NN.\nPlease let me know if you're interested.", "votes": null }, { "id": "596578", "postDate": "08/10/2019 22:22:21", "content": "Are you guys working on a GNN model by chance?
", "rawMarkdown": "Are you guys working on a GNN model by chance?", "votes": null }, { "id": "596582", "postDate": "08/10/2019 22:49:38", "content": "current score is =LGBM, except we are using MPNN submission for 3JHC and 2JHC since haven't finished running those LGBMmodels yet. i have a FFNN stack set up that marginally improves score when combining results from MPNN and LGBM
", "rawMarkdown": "current score is =LGBM, except we are using MPNN submission for 3JHC and 2JHC since haven't finished running those LGBMmodels yet. i have a FFNN stack set up that marginally improves score when combining results from MPNN and LGBM", "votes": null }, { "id": "597208", "postDate": "08/12/2019 02:31:18", "content": "I am also looking for a team :)
", "rawMarkdown": "I am also looking for a team :)", "votes": null }, { "id": "597246", "postDate": "08/12/2019 03:55:00", "content": "hi @lpachuong we can be a good fit, connect with me at satyamsharma211999@gmail.com if you're interested.
", "rawMarkdown": "hi @lpachuong we can be a good fit, connect with me at satyamsharma211999@gmail.com if you're interested.", "votes": null } ], "comments": [ { "id": 585636, "author_name": "returnofsputnik", "author_url": "", "post_date": "07/27/2019 19:16:05", "content": "Very interested to work with you. currently have 3jhh at 0.12 MAE and 1jhc at 0.487 MAE. Training 3jhn and I at 0.0904 MAE right now. I havent tried to submit on LB for awhile though, so my current ranking does not reflect the updated CV's.
\n\ni private messaged you if you are interested
", "votes": null, "replies": [ { "id": 596578, "author_name": "scaomath", "author_url": "", "post_date": "08/10/2019 22:22:21", "content": "Are you guys working on a GNN model by chance?
", "votes": null, "replies": [] }, { "id": 596582, "author_name": "returnofsputnik", "author_url": "", "post_date": "08/10/2019 22:49:38", "content": "current score is =LGBM, except we are using MPNN submission for 3JHC and 2JHC since haven't finished running those LGBMmodels yet. i have a FFNN stack set up that marginally improves score when combining results from MPNN and LGBM
", "votes": null, "replies": [] } ] }, { "id": 585680, "author_name": "stanislavblinov", "author_url": "", "post_date": "07/27/2019 20:56:43", "content": "Also interested.\nMy model is lgbm, but I didn't work at this competition at all for 3 weeks already (I'm on vacation with no access to my computer). My CV (it's a bit old, but there were no significant changes):\ntype | log | mae |\n1JHC | -0.825 | 0.573 |\n1JHN | -1.650 | 0.296 |\n2JHC | -1.273 | 0.308 |\n2JHH | -2.138 | 0.151 |\n2JHN | -1.992 | 0.193 |\n3JHC | -1.841 | 0.433 |\n3JHH | -1.816 | 0.212 |\n3JHN | -2.32 | 0.157 |\nI'll be able to start participating again in 3-4 days (~1 Aug.).\nMy plan is to create features anew because for a moment I have a lot of trash in it and then to start building a NN.\nPlease let me know if you're interested.
", "votes": null, "replies": [ { "id": 597208, "author_name": "lpachuong", "author_url": "", "post_date": "08/12/2019 02:31:18", "content": "I am also looking for a team :)
", "votes": null, "replies": [] }, { "id": 597246, "author_name": "satyamsharma", "author_url": "", "post_date": "08/12/2019 03:55:00", "content": "hi @lpachuong we can be a good fit, connect with me at satyamsharma211999@gmail.com if you're interested.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "585610": "My score is not (yet) brilliant and I should explain my work so far. I've been too busy to give this competition my full attention, but over the summer my workload has decreased and I now have around 8 hours a day to focus on it - essentially a full-time job. \n\nMy model currently uses only structural and mulliken charge features - none of Giba's features regarding Yukawa or Coulomb fields, and none of the other train datasets or QM9. I have engineered my own features to be invariant of the various symmetries present in molecules - most people seem to be using features that are dependant on `atom_index` numbers and x, y & z coordinates which I think are prone to identifying misleading correlations. I have reason to believe my features are strong, and my low score may be due to train/test feature distributions being skewed. I'm reducing their number accordingly. With 2JHH for example I can get a CV score of -2 using only 29 features.\n\nI have used my knowledge of organic chemistry to engineer features that are consistent with factors that affect their molecular properties, including but not limited to, ring strain, dihedral angles, electronegative substituents and pi/sigma characteristics of organic bond types. I have done quite a bit of research on my own to to this end, and certainly have amassed some reading material that would be of interest to anyone! I have tried to make features that I didn't think would occur to most people, and I have been fortunate that many of these have proven to be important features for my models. I think I have a lot to offer any team.\n\nAs for myself, I have a BSc in Physics and an MSc in an unrelated computing subject. I have worked previously with NMR data and while we are not allowed to use QM calculations for this competition my understanding may help. I have been studying ML for around 18 months and have broad experience with NNs and GBDT models. On that note, I feel that graphical neural networks are the most appropriate way forward with molecular data. The features I have engineered lend themselves well to this framework since they are rotationally invariant.\n\nI am hardworking, competitive and driven by curiosity - for what it's worth, I've been told that I'm a pleasure to work with! If any team is interested, please message me over this weekend. I look forward to contributing.", "585636": "Very interested to work with you. currently have 3jhh at 0.12 MAE and 1jhc at 0.487 MAE. Training 3jhn and I at 0.0904 MAE right now. I havent tried to submit on LB for awhile though, so my current ranking does not reflect the updated CV's.\n\ni private messaged you if you are interested", "585680": "Also interested.\nMy model is lgbm, but I didn't work at this competition at all for 3 weeks already (I'm on vacation with no access to my computer). My CV (it's a bit old, but there were no significant changes):\ntype | log | mae |\n1JHC | -0.825 | 0.573 |\n1JHN | -1.650 | 0.296 |\n2JHC | -1.273 | 0.308 |\n2JHH | -2.138 | 0.151 |\n2JHN | -1.992 | 0.193 |\n3JHC | -1.841 | 0.433 |\n3JHH | -1.816 | 0.212 |\n3JHN | -2.32 | 0.157 |\nI'll be able to start participating again in 3-4 days (~1 Aug.).\nMy plan is to create features anew because for a moment I have a lot of trash in it and then to start building a NN.\nPlease let me know if you're interested.", "596578": "Are you guys working on a GNN model by chance?", "596582": "current score is =LGBM, except we are using MPNN submission for 3JHC and 2JHC since haven't finished running those LGBMmodels yet. i have a FFNN stack set up that marginally improves score when combining results from MPNN and LGBM", "597208": "I am also looking for a team :)", "597246": "hi @lpachuong we can be a good fit, connect with me at satyamsharma211999@gmail.com if you're interested." }, "source": "meta" }