![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F172860%2Fe6da3bfe5c2843a00a6f445008dd9230%2Flofo_giba_features.png?generation=1562638851783929&alt=media\")
I have run the LOFO importance algorithm on Giba's features (one single run, 10% of the data, for all the bond types). \nHere is the importance plot:\n
Based on this plot, here is the list of features to dismiss:
\n\n\n['vander_F.x',\n 'atom_index_1.1',\n 'mean_molecule_atom_1_dist_xyz',\n 'adH4',\n 'mean_molecule_atom_0_dist_xyz',\n 'inv_distPE',\n 'coulomb_C.y',\n 'adC4',\n 'NC',\n 'yukawa_F.x',\n 'distN0',\n 'distC1',\n 'vander_C.y',\n 'coulomb_N.x']\n\nIt might be better to do a run for each bond type?\nI am new to this technique and any comment is welcome. Thanks. 😸
I have run the LOFO importance algorithm on Giba's features (one single run, 10% of the data, for all the bond types). \nHere is the importance plot:\n
Based on this plot, here is the list of features to dismiss:
\n\n\n['vander_F.x',\n 'atom_index_1.1',\n 'mean_molecule_atom_1_dist_xyz',\n 'adH4',\n 'mean_molecule_atom_0_dist_xyz',\n 'inv_distPE',\n 'coulomb_C.y',\n 'adC4',\n 'NC',\n 'yukawa_F.x',\n 'distN0',\n 'distC1',\n 'vander_C.y',\n 'coulomb_N.x']\n\nIt might be better to do a run for each bond type?\nI am new to this technique and any comment is welcome. Thanks. 😸
As promised, the LOFO importance plots per bond type. The notebook is coming soon. 👀
", "rawMarkdown": "As promised, the LOFO importance plots per bond type. The notebook is coming soon. 👀 ", "votes": 8, "replies": [ { "id": 571763, "postDate": "2019-07-10T04:25:47.053Z", "content": "Looking at the error bars on 1JHN, I think it might be better to use 100% of the data rather than a 10% sample. Since there's only a small number of these, it should still be fairly quick to calculate
", "rawMarkdown": "Looking at the error bars on 1JHN, I think it might be better to use 100% of the data rather than a 10% sample. Since there's only a small number of these, it should still be fairly quick to calculate", "votes": 1 }, { "id": 571813, "postDate": "2019-07-10T05:41:57.963Z", "content": "In this new experiment, I have used all the data for each bond type. Here is the link to the notebook, there is maybe a bug, please let me know if you spot any.
", "rawMarkdown": "In this new experiment, I have used all the data for each bond type. Here is the link to the [notebook](https://www.kaggle.com/yassinealouini/lofo-importance-on-giba-s-features?scriptVersionId=16970828), there is maybe a bug, please let me know if you spot any. \n", "votes": 1 } ] }, { "id": 571014, "postDate": "2019-07-09T03:56:23.127Z", "content": "Features importance are different for each bond type. So yes, you must run for each bond separately.
", "rawMarkdown": "Features importance are different for each bond type. So yes, you must run for each bond separately.", "votes": 4, "replies": [ { "id": 571070, "postDate": "2019-07-09T06:01:49.337Z", "content": "That was my intuition as well. I will re-run it and post the new results once it is done. 👌
", "rawMarkdown": "That was my intuition as well. I will re-run it and post the new results once it is done. 👌 ", "votes": 2 } ] }, { "id": 572036, "postDate": "2019-07-10T11:26:58.893Z", "content": "I have noticed that running the same code (here and locally on my computer) gives different results.\nWhat options do I have to make these results more stable and robust? Maybe averaging over many runs with different seeds?\nAny help is welcome. :)
", "rawMarkdown": "I have noticed that running the same code (here and locally on my computer) gives different results.\nWhat options do I have to make these results more stable and robust? Maybe averaging over many runs with different seeds?\nAny help is welcome. :)", "votes": 1, "replies": [ { "id": 572074, "postDate": "2019-07-10T12:18:55.067Z", "content": "CV = KFold(n_splits=3, shuffle=True)
try giving random_state as parameter.
\n\nAlso you can get better results if you provide a LGB model with Giba's parameters, because now it is using just a default model.
", "rawMarkdown": "`CV = KFold(n_splits=3, shuffle=True)`\n\ntry giving random_state as parameter.\n\nAlso you can get better results if you provide a LGB model with Giba's parameters, because now it is using just a default model." } ] }, { "id": 571620, "postDate": "2019-07-09T20:50:02.913Z", "content": "Thanks! Waiting for the kernel 👀
", "rawMarkdown": "Thanks! Waiting for the kernel 👀 ", "votes": 1, "replies": [ { "id": 572037, "postDate": "2019-07-10T11:27:43.353Z", "content": "Here you go: https://www.kaggle.com/yassinealouini/lofo-importance-on-giba-s-features. Enjoy. :)
", "rawMarkdown": "Here you go: https://www.kaggle.com/yassinealouini/lofo-importance-on-giba-s-features. Enjoy. :)", "votes": 1 } ] }, { "id": 570989, "postDate": "2019-07-09T02:30:33.277Z", "content": "Thank you so much @yassinealouini ! and yeah, I think it'd be better to run LOFO for each bond type 👍
", "rawMarkdown": "Thank you so much @yassinealouini ! and yeah, I think it'd be better to run LOFO for each bond type 👍 ", "votes": 1 }, { "id": 571099, "postDate": "2019-07-09T06:55:26.090Z", "content": "Thank you for this, I've been scratching my head for a while around how best to reduce my features. This would make a great kernel!
", "rawMarkdown": "Thank you for this, I've been scratching my head for a while around how best to reduce my features. This would make a great kernel!", "votes": 2, "replies": [ { "id": 571188, "postDate": "2019-07-09T09:24:39.920Z", "content": "I am planning to make a notebook before the end of the week. Let me know if you are faster. 🙃
", "rawMarkdown": "I am planning to make a notebook before the end of the week. Let me know if you are faster. 🙃 ", "votes": 1 } ] }, { "id": 580797, "postDate": "2019-07-20T19:33:19.453Z", "content": "Yassine
\n\nForked you kernel - with only minor change and learning to add my own data I got it to run.
\n\nVery nice work.
\n\nNow - the key question - DO YOU BELIEVE the results?
", "rawMarkdown": "Yassine\n\nForked you kernel - with only minor change and learning to add my own data I got it to run. \n\nVery nice work.\n\nNow - the key question - DO YOU BELIEVE the results?", "replies": [ { "id": 580801, "postDate": "2019-07-20T19:38:16.150Z", "content": "Someone believed and got 5th position on Microsoft Malware Competition:)\nhttps://github.com/khas-ccip/kaggle_malware_solution_khas_ccip/blob/1b1b41b22ceee3d70b81d07f4ac7a23100986493/prepare_data.py
", "rawMarkdown": "Someone believed and got 5th position on Microsoft Malware Competition:)\nhttps://github.com/khas-ccip/kaggle_malware_solution_khas_ccip/blob/1b1b41b22ceee3d70b81d07f4ac7a23100986493/prepare_data.py", "votes": -1 }, { "id": 580854, "postDate": "2019-07-20T22:38:56.660Z", "content": "Thanks - kind of looks like another believer got 10th place in a different competition.
", "rawMarkdown": "Thanks - kind of looks like another believer got 10th place in a different competition.", "votes": 1 }, { "id": 580988, "postDate": "2019-07-21T07:59:02.830Z", "content": "In order to not mislead anyone: I don't use LOFO in this competition. I usually use it in structured data competitions.
", "rawMarkdown": "In order to not mislead anyone: I don't use LOFO in this competition. I usually use it in structured data competitions.", "votes": 2 }, { "id": 580992, "postDate": "2019-07-21T08:05:21.203Z", "content": "Ahmet, I much prefer when you say what you use, rather than what you don't use ;)
", "rawMarkdown": "Ahmet, I much prefer when you say what you use, rather than what you don't use ;)", "votes": 3 }, { "id": 580993, "postDate": "2019-07-21T08:05:47.867Z", "content": "I guess we will know about your features selection method after the end of this competition. Thanks for sharing this information and good luck. 👍
", "rawMarkdown": "I guess we will know about your features selection method after the end of this competition. Thanks for sharing this information and good luck. 👍 ", "votes": 2 }, { "id": 581417, "postDate": "2019-07-21T23:19:59.300Z", "content": "Thanks - it was rude of me to make that comment - Sry
", "rawMarkdown": "Thanks - it was rude of me to make that comment - Sry" } ] }, { "id": 580825, "postDate": "2019-07-20T20:30:01.370Z", "rawMarkdown": "", "isDeleted": true }, { "id": 572131, "postDate": "2019-07-10T14:01:32.063Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 571597, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2019-07-09T19:50:15.303000", "content": "As promised, the LOFO importance plots per bond type. The notebook is coming soon. 👀
", "votes": 8, "replies": [ { "id": 571763, "author_name": "datasaurus", "author_url": "", "post_date": "2019-07-10T04:25:47.053000", "content": "Looking at the error bars on 1JHN, I think it might be better to use 100% of the data rather than a 10% sample. Since there's only a small number of these, it should still be fairly quick to calculate
", "votes": 1, "replies": [] }, { "id": 571813, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2019-07-10T05:41:57.963000", "content": "In this new experiment, I have used all the data for each bond type. Here is the link to the notebook, there is maybe a bug, please let me know if you spot any.
", "votes": 1, "replies": [] } ] }, { "id": 571014, "author_name": "Giba", "author_url": "", "post_date": "2019-07-09T03:56:23.127000", "content": "Features importance are different for each bond type. So yes, you must run for each bond separately.
", "votes": 4, "replies": [ { "id": 571070, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2019-07-09T06:01:49.337000", "content": "That was my intuition as well. I will re-run it and post the new results once it is done. 👌
", "votes": 2, "replies": [] } ] }, { "id": 572036, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2019-07-10T11:26:58.893000", "content": "I have noticed that running the same code (here and locally on my computer) gives different results.\nWhat options do I have to make these results more stable and robust? Maybe averaging over many runs with different seeds?\nAny help is welcome. :)
", "votes": 1, "replies": [ { "id": 572074, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2019-07-10T12:18:55.067000", "content": "CV = KFold(n_splits=3, shuffle=True)
try giving random_state as parameter.
\n\nAlso you can get better results if you provide a LGB model with Giba's parameters, because now it is using just a default model.
", "votes": 0, "replies": [] } ] }, { "id": 571620, "author_name": "silverstone", "author_url": "", "post_date": "2019-07-09T20:50:02.913000", "content": "Thanks! Waiting for the kernel 👀
", "votes": 1, "replies": [ { "id": 572037, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2019-07-10T11:27:43.353000", "content": "Here you go: https://www.kaggle.com/yassinealouini/lofo-importance-on-giba-s-features. Enjoy. :)
", "votes": 1, "replies": [] } ] }, { "id": 570989, "author_name": "Nanashi", "author_url": "", "post_date": "2019-07-09T02:30:33.277000", "content": "Thank you so much @yassinealouini ! and yeah, I think it'd be better to run LOFO for each bond type 👍
", "votes": 1, "replies": [] }, { "id": 571099, "author_name": "datasaurus", "author_url": "", "post_date": "2019-07-09T06:55:26.090000", "content": "Thank you for this, I've been scratching my head for a while around how best to reduce my features. This would make a great kernel!
", "votes": 2, "replies": [ { "id": 571188, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2019-07-09T09:24:39.920000", "content": "I am planning to make a notebook before the end of the week. Let me know if you are faster. 🙃
", "votes": 1, "replies": [] } ] }, { "id": 580797, "author_name": "PC Jimmmy", "author_url": "", "post_date": "2019-07-20T19:33:19.453000", "content": "Yassine
\n\nForked you kernel - with only minor change and learning to add my own data I got it to run.
\n\nVery nice work.
\n\nNow - the key question - DO YOU BELIEVE the results?
", "votes": 0, "replies": [ { "id": 580801, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2019-07-20T19:38:16.150000", "content": "Someone believed and got 5th position on Microsoft Malware Competition:)\nhttps://github.com/khas-ccip/kaggle_malware_solution_khas_ccip/blob/1b1b41b22ceee3d70b81d07f4ac7a23100986493/prepare_data.py
", "votes": -1, "replies": [] }, { "id": 580854, "author_name": "PC Jimmmy", "author_url": "", "post_date": "2019-07-20T22:38:56.660000", "content": "Thanks - kind of looks like another believer got 10th place in a different competition.
", "votes": 1, "replies": [] }, { "id": 580988, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2019-07-21T07:59:02.830000", "content": "In order to not mislead anyone: I don't use LOFO in this competition. I usually use it in structured data competitions.
", "votes": 2, "replies": [] }, { "id": 580992, "author_name": "CPMP", "author_url": "", "post_date": "2019-07-21T08:05:21.203000", "content": "Ahmet, I much prefer when you say what you use, rather than what you don't use ;)
", "votes": 3, "replies": [] }, { "id": 580993, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2019-07-21T08:05:47.867000", "content": "I guess we will know about your features selection method after the end of this competition. Thanks for sharing this information and good luck. 👍
", "votes": 2, "replies": [] }, { "id": 581417, "author_name": "PC Jimmmy", "author_url": "", "post_date": "2019-07-21T23:19:59.300000", "content": "Thanks - it was rude of me to make that comment - Sry
", "votes": 0, "replies": [] } ] }, { "id": 580825, "author_name": "", "author_url": "", "post_date": "2019-07-20T20:30:01.370000", "content": "", "votes": 0, "replies": [] }, { "id": 572131, "author_name": "", "author_url": "", "post_date": "2019-07-10T14:01:32.063000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "570985": "I have run the [**LOFO**](https://github.com/aerdem4/lofo-importance) importance algorithm on Giba's features (one single run, 10% of the data, for all the bond types). \nHere is the importance plot:\n\n\nBased on this plot, here is the list of features to dismiss: \n\n```\n['vander_F.x',\n 'atom_index_1.1',\n 'mean_molecule_atom_1_dist_xyz',\n 'adH4',\n 'mean_molecule_atom_0_dist_xyz',\n 'inv_distPE',\n 'coulomb_C.y',\n 'adC4',\n 'NC',\n 'yukawa_F.x',\n 'distN0',\n 'distC1',\n 'vander_C.y',\n 'coulomb_N.x']\n```\nIt might be better to do a run for each bond type?\nI am new to this technique and any comment is welcome. Thanks. 😸 \n", "571597": "As promised, the LOFO importance plots per bond type. The notebook is coming soon. 👀 ", "571014": "Features importance are different for each bond type. So yes, you must run for each bond separately.", "572036": "I have noticed that running the same code (here and locally on my computer) gives different results.\nWhat options do I have to make these results more stable and robust? Maybe averaging over many runs with different seeds?\nAny help is welcome. :)", "571620": "Thanks! Waiting for the kernel 👀 ", "570989": "Thank you so much @yassinealouini ! and yeah, I think it'd be better to run LOFO for each bond type 👍 ", "571099": "Thank you for this, I've been scratching my head for a while around how best to reduce my features. This would make a great kernel!", "580797": "Yassine\n\nForked you kernel - with only minor change and learning to add my own data I got it to run. \n\nVery nice work.\n\nNow - the key question - DO YOU BELIEVE the results?", "580825": "", "572131": "" } }