Some checkpoints for people who are interested. Also would love to get inspirations from the comments.
\n(i) Sergei's starter kit can give you score ~536, outpacing half of the teams on the current leaderboard.
\n(ii) I see hundreds of teams at 545 Bronze Medal, which suggests there may be already some shared notebook you can directly use to get here.
\n(iii) To get 569, I figured out a way to extend the mean estimation. This score can get you #38 on the current leaderboard.
\n(iv) To get 602, I figured out a way to estimate sigma. This score can get you at #25 on the current leaderboard.
\n(v) A second method to improve sigma leads me to 619, ranked top20 on current leaderboard.
Some checkpoints for people who are interested. Also would love to get inspirations from the comments.
\n(i) Sergei's starter kit can give you score ~536, outpacing half of the teams on the current leaderboard.
\n(ii) I see hundreds of teams at 545 Bronze Medal, which suggests there may be already some shared notebook you can directly use to get here.
\n(iii) To get 569, I figured out a way to extend the mean estimation. This score can get you #38 on the current leaderboard.
\n(iv) To get 602, I figured out a way to estimate sigma. This score can get you at #25 on the current leaderboard.
\n(v) A second method to improve sigma leads me to 619, ranked top20 on current leaderboard.
\n\n(iv) To get 602, I figured out a way to estimate sigma. This score can get you at #25 on the current leaderboard.
\n
My local CV by cnn can approach 0.65, and sigma is my bottleneck, sad. However, Whether sigma estimation to the private test set is also questionable.
", "rawMarkdown": ">(iv) To get 602, I figured out a way to estimate sigma. This score can get you at #25 on the current leaderboard.\n\nMy local CV by cnn can approach 0.65, and sigma is my bottleneck, sad. However, Whether sigma estimation to the private test set is also questionable.\n", "votes": 3, "replies": [ { "id": 3016797, "postDate": "2024-10-14T07:09:23.487Z", "content": "My local CV also have a great score, with sigma ~ 4e-5, but somehow the sigma in test dataset is ~ 1e-4
", "rawMarkdown": "My local CV also have a great score, with sigma ~ 4e-5, but somehow the sigma in test dataset is ~ 1e-4", "votes": 3, "replies": [ { "id": 3030704, "postDate": "2024-10-28T19:32:23.870Z", "content": "so how do you adjust for test data sigma
", "rawMarkdown": "so how do you adjust for test data sigma" } ] } ] }, { "id": 3020935, "postDate": "2024-10-18T04:22:49.477Z", "content": "Thanks for sharing. Do you mind if I ask what is your mse? Because the max score is highly dependent on what your mse is. Mine have a mse of 91 [ppm] and score=566, cheers.
", "rawMarkdown": "Thanks for sharing. Do you mind if I ask what is your mse? Because the max score is highly dependent on what your mse is. Mine have a mse of 91 [ppm] and score=566, cheers.", "votes": 4, "replies": [ { "id": 3029146, "postDate": "2024-10-27T00:22:22.953Z", "content": "mine is 70 ppm
", "rawMarkdown": "mine is 70 ppm", "votes": 1 } ] }, { "id": 3027507, "postDate": "2024-10-24T23:54:34.207Z", "content": "How did you extend the mean, using a Ridge Regression or down the NN path? I am struggling with overfitting, unfortunately. My RMSE is just a bit over 90, yet any method that gets me below that overfits.
", "rawMarkdown": "How did you extend the mean, using a Ridge Regression or down the NN path? I am struggling with overfitting, unfortunately. My RMSE is just a bit over 90, yet any method that gets me below that overfits.", "votes": 1 }, { "id": 3016483, "postDate": "2024-10-13T19:58:53.523Z", "content": "Nice! I have one question - did you use Sergei's solution as a base, or did you use some completely different approach?
", "rawMarkdown": "Nice! I have one question - did you use Sergei's solution as a base, or did you use some completely different approach? ", "votes": 2, "replies": [ { "id": 3016717, "postDate": "2024-10-14T05:02:00.537Z", "content": "Yes, Sergei's solution.
", "rawMarkdown": "Yes, Sergei's solution.", "votes": 2 } ] }, { "id": 3017973, "postDate": "2024-10-15T12:25:21.973Z", "content": "Hello :) By \"extend the mean\", you mean predict more wavelength than just a single mean value ?
", "rawMarkdown": "Hello :) By \"extend the mean\", you mean predict more wavelength than just a single mean value ?", "replies": [ { "id": 3018573, "postDate": "2024-10-15T21:18:23.180Z", "content": "Yes, that's what I meant.
", "rawMarkdown": "Yes, that's what I meant. ", "votes": 1, "replies": [ { "id": 3022626, "postDate": "2024-10-19T19:06:47.697Z", "content": "For predicting the mean value of transit depth (same for all wavelengths), was there any need to use model with parameters?
\nthank you Lien
Hi, I think your prediction looks ok on training set. It may be just a problem of generalization. And not surprisingly the more parameters you have in the model, the higher change it gonna suffer overfitting.
", "rawMarkdown": "Hi, I think your prediction looks ok on training set. It may be just a problem of generalization. And not surprisingly the more parameters you have in the model, the higher change it gonna suffer overfitting. " } ] } ] }, { "id": 3025956, "postDate": "2024-10-23T09:53:57.107Z", "content": "I really look forward to see how you handled the sigma estimate !
\nOn my side, the estimates I obtain from the training set do not match at all with what I could infer from the LB …
", "rawMarkdown": "I really look forward to see how you handled the sigma estimate ! \n\nOn my side, the estimates I obtain from the training set do not match at all with what I could infer from the LB ... ", "replies": [ { "id": 3025994, "postDate": "2024-10-23T10:51:42.520Z", "content": "Due to the difference between train and test set distributions, test set sigma will be higher than train estimate in any case.
", "rawMarkdown": "Due to the difference between train and test set distributions, test set sigma will be higher than train estimate in any case.", "votes": 1 } ] }, { "id": 3018911, "postDate": "2024-10-16T05:56:18.953Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3016768, "author_name": "Timmy Juicehouse", "author_url": "", "post_date": "2024-10-14T06:19:35.310000", "content": "\n\n(iv) To get 602, I figured out a way to estimate sigma. This score can get you at #25 on the current leaderboard.
\n
My local CV by cnn can approach 0.65, and sigma is my bottleneck, sad. However, Whether sigma estimation to the private test set is also questionable.
", "votes": 3, "replies": [ { "id": 3016797, "author_name": "ChingYinNg", "author_url": "", "post_date": "2024-10-14T07:09:23.487000", "content": "My local CV also have a great score, with sigma ~ 4e-5, but somehow the sigma in test dataset is ~ 1e-4
", "votes": 3, "replies": [ { "id": 3030704, "author_name": "Ajay Varghese", "author_url": "", "post_date": "2024-10-28T19:32:23.870000", "content": "so how do you adjust for test data sigma
", "votes": 0, "replies": [] } ] } ] }, { "id": 3020935, "author_name": "Wei-Kuo Li", "author_url": "", "post_date": "2024-10-18T04:22:49.477000", "content": "Thanks for sharing. Do you mind if I ask what is your mse? Because the max score is highly dependent on what your mse is. Mine have a mse of 91 [ppm] and score=566, cheers.
", "votes": 4, "replies": [ { "id": 3029146, "author_name": "Lien", "author_url": "", "post_date": "2024-10-27T00:22:22.953000", "content": "mine is 70 ppm
", "votes": 1, "replies": [] } ] }, { "id": 3027507, "author_name": "Andrei Zamfir", "author_url": "", "post_date": "2024-10-24T23:54:34.207000", "content": "How did you extend the mean, using a Ridge Regression or down the NN path? I am struggling with overfitting, unfortunately. My RMSE is just a bit over 90, yet any method that gets me below that overfits.
", "votes": 1, "replies": [] }, { "id": 3016483, "author_name": "Araik Tamazian", "author_url": "", "post_date": "2024-10-13T19:58:53.523000", "content": "Nice! I have one question - did you use Sergei's solution as a base, or did you use some completely different approach?
", "votes": 2, "replies": [ { "id": 3016717, "author_name": "Lien", "author_url": "", "post_date": "2024-10-14T05:02:00.537000", "content": "Yes, Sergei's solution.
", "votes": 2, "replies": [] } ] }, { "id": 3017973, "author_name": "dafram2r", "author_url": "", "post_date": "2024-10-15T12:25:21.973000", "content": "Hello :) By \"extend the mean\", you mean predict more wavelength than just a single mean value ?
", "votes": 0, "replies": [ { "id": 3018573, "author_name": "Lien", "author_url": "", "post_date": "2024-10-15T21:18:23.180000", "content": "Yes, that's what I meant.
", "votes": 1, "replies": [ { "id": 3022626, "author_name": "highDopamine", "author_url": "", "post_date": "2024-10-19T19:06:47.697000", "content": "For predicting the mean value of transit depth (same for all wavelengths), was there any need to use model with parameters?
\nthank you Lien
Hi, I think your prediction looks ok on training set. It may be just a problem of generalization. And not surprisingly the more parameters you have in the model, the higher change it gonna suffer overfitting.
", "votes": 0, "replies": [] } ] } ] }, { "id": 3025956, "author_name": "FabienDaniel", "author_url": "", "post_date": "2024-10-23T09:53:57.107000", "content": "I really look forward to see how you handled the sigma estimate !
\nOn my side, the estimates I obtain from the training set do not match at all with what I could infer from the LB …
", "votes": 0, "replies": [ { "id": 3025994, "author_name": "Araik Tamazian", "author_url": "", "post_date": "2024-10-23T10:51:42.520000", "content": "Due to the difference between train and test set distributions, test set sigma will be higher than train estimate in any case.
", "votes": 1, "replies": [] } ] }, { "id": 3018911, "author_name": "", "author_url": "", "post_date": "2024-10-16T05:56:18.953000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3016421": "Some checkpoints for people who are interested. Also would love to get inspirations from the comments.\n(i) Sergei's starter kit can give you score ~536, outpacing half of the teams on the current leaderboard.\n(ii) I see hundreds of teams at 545 Bronze Medal, which suggests there may be already some shared notebook you can directly use to get here.\n(iii) To get 569, I figured out a way to extend the mean estimation. This score can get you #38 on the current leaderboard.\n(iv) To get 602, I figured out a way to estimate sigma. This score can get you at #25 on the current leaderboard.\n(v) A second method to improve sigma leads me to 619, ranked top20 on current leaderboard.\n\n", "3016768": ">(iv) To get 602, I figured out a way to estimate sigma. This score can get you at #25 on the current leaderboard.\n\nMy local CV by cnn can approach 0.65, and sigma is my bottleneck, sad. However, Whether sigma estimation to the private test set is also questionable.\n", "3020935": "Thanks for sharing. Do you mind if I ask what is your mse? Because the max score is highly dependent on what your mse is. Mine have a mse of 91 [ppm] and score=566, cheers.", "3027507": "How did you extend the mean, using a Ridge Regression or down the NN path? I am struggling with overfitting, unfortunately. My RMSE is just a bit over 90, yet any method that gets me below that overfits.", "3016483": "Nice! I have one question - did you use Sergei's solution as a base, or did you use some completely different approach? ", "3017973": "Hello :) By \"extend the mean\", you mean predict more wavelength than just a single mean value ?", "3025956": "I really look forward to see how you handled the sigma estimate ! \n\nOn my side, the estimates I obtain from the training set do not match at all with what I could infer from the LB ... ", "3018911": "" } }