When there are multiple planets available I currently determine the target value for each planet and then I average them.
\nThis gives a modest performance boost.
\nAre there better ways to use the results of multiple planets?
", "messages": [ { "id": 3273725, "postDate": "2025-08-23T09:19:23.660Z", "content": "When there are multiple planets available I currently determine the target value for each planet and then I average them.
\nThis gives a modest performance boost.
\nAre there better ways to use the results of multiple planets?
", "rawMarkdown": "When there are multiple planets available I currently determine the target value for each planet and then I average them.\n\nThis gives a modest performance boost.\n\nAre there better ways to use the results of multiple planets?", "votes": 6 }, { "id": 3274062, "postDate": "2025-08-23T23:05:19.640Z", "content": "Hi, how do you guys determine which planets have multiple measurements?
", "rawMarkdown": "Hi, how do you guys determine which planets have multiple measurements?", "votes": 1, "replies": [ { "id": 3274137, "postDate": "2025-08-24T06:35:20.990Z", "content": "Check if 1 file exists with _1. Than load and proces all files.
", "rawMarkdown": "Check if 1 file exists with _1. Than load and proces all files.", "votes": 2 } ] }, { "id": 3273742, "postDate": "2025-08-23T10:24:55.780Z", "content": "Just a quick note, it is the same planet, but multiple measurements.
\nAveraging is the most straight forward solution. But you may want to take into account the uncertainty in the measurements and weight accordingly.
@ilu000 Thanks yes twisted it around 🤣 1 planet and multiple measurements.
\ninteresting so a weighted average for both the predictions and the uncertainty. Will give that a try indeed.
\nThanks!
My LB does not change when including second transits and CV doesn't improve either, quite strange.
", "rawMarkdown": "My LB does not change when including second transits and CV doesn't improve either, quite strange.", "replies": [ { "id": 3287623, "postDate": "2025-09-11T23:06:49.077Z", "content": "I think the issue is if hidden set is similar to train set (and I have reason to believe, but can't confirm), the second transits only make up 10% of the dataset. So unless they improve your score by a lot for those examples, they are going make a very little impact.
", "rawMarkdown": "I think the issue is if hidden set is similar to train set (and I have reason to believe, but can't confirm), the second transits only make up 10% of the dataset. So unless they improve your score by a lot for those examples, they are going make a very little impact.", "replies": [ { "id": 3287715, "postDate": "2025-09-12T05:08:30.800Z", "content": "That's true, but I still find it strange that my LB numbers are exactly the same, would at least expect a small perturbation.
", "rawMarkdown": "That's true, but I still find it strange that my LB numbers are exactly the same, would at least expect a small perturbation." } ] } ] }, { "id": 3275172, "postDate": "2025-08-26T04:00:05.593Z", "content": "Currently select the prediction with lower std. But I am facing overfitting problems. Do you have tips for breaking a 0 score for NN methods with 1d input?
", "rawMarkdown": "Currently select the prediction with lower std. But I am facing overfitting problems. Do you have tips for breaking a 0 score for NN methods with 1d input?", "replies": [ { "id": 3275214, "postDate": "2025-08-26T06:16:27.400Z", "content": "A score of 0 likely has little to do with which model you used, as the hosts updated the training dataset. You need to adjust your strategy for detecting transits. If you had performed EDA, you would have noticed that data in 2025 is much more complex compared to the last year—especially the transit trajectories, which vary dramatically, with many extremely unusual cases.
", "rawMarkdown": "A score of 0 likely has little to do with which model you used, as the hosts updated the training dataset. You need to adjust your strategy for detecting transits. If you had performed EDA, you would have noticed that data in 2025 is much more complex compared to the last year—especially the transit trajectories, which vary dramatically, with many extremely unusual cases.", "votes": 1 }, { "id": 3275215, "postDate": "2025-08-26T06:18:31.800Z", "content": "in a word, probably you found wrong transits.
", "rawMarkdown": "in a word, probably you found wrong transits.", "replies": [ { "id": 3275391, "postDate": "2025-08-26T12:29:12.297Z", "content": "Good catch, I am still out of the right track
", "rawMarkdown": "Good catch, I am still out of the right track" } ] } ] } ], "comments": [ { "id": 3274062, "author_name": "noobanot", "author_url": "", "post_date": "2025-08-23T23:05:19.640000", "content": "Hi, how do you guys determine which planets have multiple measurements?
", "votes": 1, "replies": [ { "id": 3274137, "author_name": "Robin Smits", "author_url": "", "post_date": "2025-08-24T06:35:20.990000", "content": "Check if 1 file exists with _1. Than load and proces all files.
", "votes": 2, "replies": [] } ] }, { "id": 3273742, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2025-08-23T10:24:55.780000", "content": "Just a quick note, it is the same planet, but multiple measurements.
\nAveraging is the most straight forward solution. But you may want to take into account the uncertainty in the measurements and weight accordingly.
@ilu000 Thanks yes twisted it around 🤣 1 planet and multiple measurements.
\ninteresting so a weighted average for both the predictions and the uncertainty. Will give that a try indeed.
\nThanks!
My LB does not change when including second transits and CV doesn't improve either, quite strange.
", "votes": 0, "replies": [ { "id": 3287623, "author_name": "JungleBeastDS", "author_url": "", "post_date": "2025-09-11T23:06:49.077000", "content": "I think the issue is if hidden set is similar to train set (and I have reason to believe, but can't confirm), the second transits only make up 10% of the dataset. So unless they improve your score by a lot for those examples, they are going make a very little impact.
", "votes": 0, "replies": [ { "id": 3287715, "author_name": "sroger", "author_url": "", "post_date": "2025-09-12T05:08:30.800000", "content": "That's true, but I still find it strange that my LB numbers are exactly the same, would at least expect a small perturbation.
", "votes": 0, "replies": [] } ] } ] }, { "id": 3275172, "author_name": "shanzhong8", "author_url": "", "post_date": "2025-08-26T04:00:05.593000", "content": "Currently select the prediction with lower std. But I am facing overfitting problems. Do you have tips for breaking a 0 score for NN methods with 1d input?
", "votes": 0, "replies": [ { "id": 3275214, "author_name": "Timmy Juicehouse", "author_url": "", "post_date": "2025-08-26T06:16:27.400000", "content": "A score of 0 likely has little to do with which model you used, as the hosts updated the training dataset. You need to adjust your strategy for detecting transits. If you had performed EDA, you would have noticed that data in 2025 is much more complex compared to the last year—especially the transit trajectories, which vary dramatically, with many extremely unusual cases.
", "votes": 1, "replies": [] }, { "id": 3275215, "author_name": "Timmy Juicehouse", "author_url": "", "post_date": "2025-08-26T06:18:31.800000", "content": "in a word, probably you found wrong transits.
", "votes": 0, "replies": [ { "id": 3275391, "author_name": "shanzhong8", "author_url": "", "post_date": "2025-08-26T12:29:12.297000", "content": "Good catch, I am still out of the right track
", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3273725": "When there are multiple planets available I currently determine the target value for each planet and then I average them.\n\nThis gives a modest performance boost.\n\nAre there better ways to use the results of multiple planets?", "3274062": "Hi, how do you guys determine which planets have multiple measurements?", "3273742": "Just a quick note, it is the same planet, but multiple measurements.\nAveraging is the most straight forward solution. But you may want to take into account the uncertainty in the measurements and weight accordingly.", "3287468": "My LB does not change when including second transits and CV doesn't improve either, quite strange.", "3275172": "Currently select the prediction with lower std. But I am facing overfitting problems. Do you have tips for breaking a 0 score for NN methods with 1d input?" } }