![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11676771%2Fc4f0cb08fc61cec421a53b1f9d45dc04%2F3d4718a8-9655-4ca9-ac8b-b30b25c8881b.png?generation=1758790994743409&alt=media\")
In this particular case my predicts were consistently 2x lower compared to the actual and I couldn't figure a way to improve this. In the end I just multiplied sigmas for planets without visible ingress/egress by some fixed multiplier found by CV.
", "messages": [ { "id": 3294047, "postDate": "2025-09-25T09:03:54.377Z", "content": "My approach is as I have described in the comment of our solution:
\n\n\nIn fact, I didn’t handle these cases directly. My approach is: when the phase detector flags this type of sample, I set the time points t1, t2, t3, t4 to the two sides (edges). The idea is that, for such samples, connecting the two endpoints on both sides to form a straight line is the most reasonable baseline. (There will still be loss, but I think this gets as close as we can.)At the same time, when fitting the baseline, use lower-order polynomials (as I have written, such as a second-order polynomial).
\n
Reducing the polynomial degree of the baseline resulted in a roughly 0.01 improvement for me.
\nfor eg. Index 496
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11676771%2F99419fb9ac1fe55d0f17f9d3c6530524%2F0ec39ab5-8149-4990-9241-cbf028670bc3.png?generation=1758791002380695&alt=media\")
Thanks! I used a somewhat similar approach using a constant (mean value) as a baseline and refining it afterwards. My problem was that sometimes it worked, sometimes it failed spectacularly and I couldn't find a heuristic that worked reasonably well for those cases. That's why in the end I resorted to increasing the sigma as a protection against catastrophic negative scores.
", "rawMarkdown": "Thanks! I used a somewhat similar approach using a constant (mean value) as a baseline and refining it afterwards. My problem was that sometimes it worked, sometimes it failed spectacularly and I couldn't find a heuristic that worked reasonably well for those cases. That's why in the end I resorted to increasing the sigma as a protection against catastrophic negative scores.", "votes": 1, "replies": [ { "id": 3294059, "postDate": "2025-09-25T09:27:45.143Z", "content": "I think it's reasonable. We had the same plan before, but in our final analysis, we found that our method basically avoided the value of 0 for all the abnormal samples in the training set. Therefore, we didn't choose this approach.
", "rawMarkdown": "I think it's reasonable. We had the same plan before, but in our final analysis, we found that our method basically avoided the value of 0 for all the abnormal samples in the training set. Therefore, we didn't choose this approach.", "votes": 1 } ] } ] }, { "id": 3294089, "postDate": "2025-09-25T10:34:10.783Z", "content": "I'll make a writeup to explain my solution, but I'll probably not get it done today, so here is a quick look at 158006264.
\nI made a physical model of one circle passing in front of another. The transit is active while the two circles are touching. The following two figures are transits for 158006264. The dashed vertical lines show the beginning, middle, and end of the transit. The red and green lines are predictions with and without the transit component. For these figures I just summed AIRS over the entire spectrum.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11013868%2Fe264c24f61f0ea8f15be8da59bbc027f%2F0_158006264_airs_sum_-12.94268257_-27.58614617_-39.66989943.png?generation=1758795772953181&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11013868%2F4d9123ec9d2653833c3e6c72eda07e47%2F1_158006264_airs_sum_-22.71863675_-34.54346617___2.50379717.png?generation=1758796106961891&alt=media\")
Interestingly, this approach allows the model to predict a full transit from only part of the data. The quality does suffer, however. My worst predictions are for planet 1843015807, and they look as follows:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11013868%2Ffc5d58230c71c9f3a35b73f2fb64d97b%2F2_1843015807_airs_sum_-121.06615705_-144.61503641__180.37032556.png?generation=1758796374789667&alt=media\")
In this particular case my predicts were consistently 2x lower compared to the actual and I couldn't figure a way to improve this. In the end I just multiplied sigmas for planets without visible ingress/egress by some fixed multiplier found by CV.
", "rawMarkdown": "In this particular case my predicts were consistently 2x lower compared to the actual and I couldn't figure a way to improve this. In the end I just multiplied sigmas for planets without visible ingress/egress by some fixed multiplier found by CV.", "votes": 2 } ], "comments": [ { "id": 3294047, "author_name": "Horikita Saku", "author_url": "", "post_date": "2025-09-25T09:03:54.377000", "content": "My approach is as I have described in the comment of our solution:
\n\n\nIn fact, I didn’t handle these cases directly. My approach is: when the phase detector flags this type of sample, I set the time points t1, t2, t3, t4 to the two sides (edges). The idea is that, for such samples, connecting the two endpoints on both sides to form a straight line is the most reasonable baseline. (There will still be loss, but I think this gets as close as we can.)At the same time, when fitting the baseline, use lower-order polynomials (as I have written, such as a second-order polynomial).
\n
Reducing the polynomial degree of the baseline resulted in a roughly 0.01 improvement for me.
\nfor eg. Index 496
\n
Thanks! I used a somewhat similar approach using a constant (mean value) as a baseline and refining it afterwards. My problem was that sometimes it worked, sometimes it failed spectacularly and I couldn't find a heuristic that worked reasonably well for those cases. That's why in the end I resorted to increasing the sigma as a protection against catastrophic negative scores.
", "votes": 1, "replies": [ { "id": 3294059, "author_name": "Horikita Saku", "author_url": "", "post_date": "2025-09-25T09:27:45.143000", "content": "I think it's reasonable. We had the same plan before, but in our final analysis, we found that our method basically avoided the value of 0 for all the abnormal samples in the training set. Therefore, we didn't choose this approach.
", "votes": 1, "replies": [] } ] } ] }, { "id": 3294089, "author_name": "Thomas Dueholm Hansen", "author_url": "", "post_date": "2025-09-25T10:34:10.783000", "content": "I'll make a writeup to explain my solution, but I'll probably not get it done today, so here is a quick look at 158006264.
\nI made a physical model of one circle passing in front of another. The transit is active while the two circles are touching. The following two figures are transits for 158006264. The dashed vertical lines show the beginning, middle, and end of the transit. The red and green lines are predictions with and without the transit component. For these figures I just summed AIRS over the entire spectrum.
\n
\n
Interestingly, this approach allows the model to predict a full transit from only part of the data. The quality does suffer, however. My worst predictions are for planet 1843015807, and they look as follows:
\n