There have been several discussions and kernels around Mulliken charges. As you know, we only have competition-supplied values for the training dataset.
\n\nI decided to take advantage of the excellent kernel:\nhttps://www.kaggle.com/asauve/v7-estimation-of-mulliken-charges-with-open-babel/notebook\ndeveloped by Alexandre Sauvé using Open Babel to provide the missing MC values for the test dataset.
\n\nI used the competition MC dataset and the Open Babel data for the test set. My model severely underfit the leaderboard - CV ~ -0.8, LB 0.5.
\n\nThere has been some speculation that the MC values provided were taken at a different temperature, which could account for this large discrepancy.
\n\nI then used the Open Babel for both training and test. My LB score of -0.8 was consistent with my CV score, I concluded that the competition-supplied Mulliken charge values are probably not very useful.
\n\nHas anyone had similar results?
", "messages": [ { "id": "556138", "postDate": "06/19/2019 22:15:55", "content": "There have been several discussions and kernels around Mulliken charges. As you know, we only have competition-supplied values for the training dataset.
\n\nI decided to take advantage of the excellent kernel:\nhttps://www.kaggle.com/asauve/v7-estimation-of-mulliken-charges-with-open-babel/notebook\ndeveloped by Alexandre Sauvé using Open Babel to provide the missing MC values for the test dataset.
\n\nI used the competition MC dataset and the Open Babel data for the test set. My model severely underfit the leaderboard - CV ~ -0.8, LB 0.5.
\n\nThere has been some speculation that the MC values provided were taken at a different temperature, which could account for this large discrepancy.
\n\nI then used the Open Babel for both training and test. My LB score of -0.8 was consistent with my CV score, I concluded that the competition-supplied Mulliken charge values are probably not very useful.
\n\nHas anyone had similar results?
", "rawMarkdown": "There have been several discussions and kernels around Mulliken charges. As you know, we only have competition-supplied values for the training dataset.\n\nI decided to take advantage of the excellent kernel:\nhttps://www.kaggle.com/asauve/v7-estimation-of-mulliken-charges-with-open-babel/notebook\ndeveloped by Alexandre Sauvé using Open Babel to provide the missing MC values for the test dataset.\n\nI used the competition MC dataset and the Open Babel data for the test set. My model severely underfit the leaderboard - CV ~ -0.8, LB 0.5.\n\nThere has been some speculation that the MC values provided were taken at a different temperature, which could account for this large discrepancy.\n\nI then used the Open Babel for both training and test. My LB score of -0.8 was consistent with my CV score, I concluded that the competition-supplied Mulliken charge values are probably not very useful.\n\nHas anyone had similar results?", "votes": null }, { "id": "556341", "postDate": "06/20/2019 06:09:00", "content": "I'm not using the supplied charges at the moment. I figured there are people out there smarter than me who have put effort into creating physics based partial charge models which will probably perform better than my predictions derived from a small-ish dataset
", "rawMarkdown": "I'm not using the supplied charges at the moment. I figured there are people out there smarter than me who have put effort into creating physics based partial charge models which will probably perform better than my predictions derived from a small-ish dataset", "votes": null } ], "comments": [ { "id": 556341, "author_name": "anjum48", "author_url": "", "post_date": "06/20/2019 06:09:00", "content": "I'm not using the supplied charges at the moment. I figured there are people out there smarter than me who have put effort into creating physics based partial charge models which will probably perform better than my predictions derived from a small-ish dataset
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "556138": "There have been several discussions and kernels around Mulliken charges. As you know, we only have competition-supplied values for the training dataset.\n\nI decided to take advantage of the excellent kernel:\nhttps://www.kaggle.com/asauve/v7-estimation-of-mulliken-charges-with-open-babel/notebook\ndeveloped by Alexandre Sauvé using Open Babel to provide the missing MC values for the test dataset.\n\nI used the competition MC dataset and the Open Babel data for the test set. My model severely underfit the leaderboard - CV ~ -0.8, LB 0.5.\n\nThere has been some speculation that the MC values provided were taken at a different temperature, which could account for this large discrepancy.\n\nI then used the Open Babel for both training and test. My LB score of -0.8 was consistent with my CV score, I concluded that the competition-supplied Mulliken charge values are probably not very useful.\n\nHas anyone had similar results?", "556341": "I'm not using the supplied charges at the moment. I figured there are people out there smarter than me who have put effort into creating physics based partial charge models which will probably perform better than my predictions derived from a small-ish dataset" }, "source": "meta" }