![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F20811297%2Fb25db0d72721d33ff2822c44bd4113a1%2Fr.png?generation=1726483662607168&alt=media\")
What is the secret to achieving the top 10 LB places?
\n
Perhaps there is some information already in this discussion.
\nWho can review everything we know about competition?
\nIs there any important post list?
Without giving anything away, my approach is along the lines of:
\nThen, the killer, reimplement everything in that pipeline to take advantage of this problem's structure to avoid bottlenecks in scipy/numpy. Otherwise, the test set would take ~200 hours to solve for.
\nI think my method will top out at about 0.63 unless I have some sort of additional insight.
", "rawMarkdown": "Without giving anything away, my approach is along the lines of:\n- Get a \"cleaner\" signal.\n- Use optimization to get a best fit for depth.\n- Get a theoretical estimate of sigma from your best fit.\n- Model the noise given some basic domain knowledge. Get a different estimate of sigma.\n- Find a consensus estimate of sigma from your two approaches.\n\nThen, the killer, reimplement everything in that pipeline to take advantage of this problem's structure to avoid bottlenecks in scipy/numpy. Otherwise, the test set would take ~200 hours to solve for.\n\nI think my method will top out at about 0.63 unless I have some sort of additional insight.", "votes": 18, "replies": [ { "id": 2990694, "postDate": "2024-09-16T16:09:58.353Z", "content": "Thanks, how much \"cleaner\" signal improves your score? Or what do you mean by \"cleaner\" signal? As for me, no denoising algorithm has done well.
", "rawMarkdown": "Thanks, how much \"cleaner\" signal improves your score? Or what do you mean by \"cleaner\" signal? As for me, no denoising algorithm has done well.", "votes": 6, "replies": [ { "id": 2991013, "postDate": "2024-09-16T23:28:26.477Z", "content": "I've tried close to a dozen approaches so far. There are small gains from some of them. The reason I put \"cleaner\" in quotes is, in some sense, simply averaging all the wavelengths provides you with a cleaner signal.
", "rawMarkdown": "I've tried close to a dozen approaches so far. There are small gains from some of them. The reason I put \"cleaner\" in quotes is, in some sense, simply averaging all the wavelengths provides you with a cleaner signal.", "votes": 6 } ] } ] }, { "id": 2990378, "postDate": "2024-09-16T10:52:39.280Z", "content": "What is the secret to achieving the top 10 LB places?
\n
Perhaps there is some information already in this discussion.
\nWho can review everything we know about competition?
\nIs there any important post list?
I'm half hoping no one will reveal any 'magic' so I can stay in top 5, half hoping that someone will reveal exactly the 'magic' that I'm missing to get to 0.65+ range 🤣
", "rawMarkdown": "I'm half hoping no one will reveal any 'magic' so I can stay in top 5, half hoping that someone will reveal exactly the 'magic' that I'm missing to get to 0.65+ range 🤣", "votes": 7 }, { "id": 2990478, "postDate": "2024-09-16T12:22:28.070Z", "content": "I'm pretty sure it has to do with a good estimation of the uncertainty part.
", "rawMarkdown": "I'm pretty sure it has to do with a good estimation of the uncertainty part.", "votes": 3, "replies": [ { "id": 2991112, "postDate": "2024-09-17T04:15:50.533Z", "content": "Yep. A very rough estimate of sigma boosted my score by 0.03
", "rawMarkdown": "Yep. A very rough estimate of sigma boosted my score by 0.03", "votes": 3 } ] }, { "id": 2990398, "postDate": "2024-09-16T11:01:46.637Z", "content": "@arabidopsisthalian we are also struggling to find the magic. at present -> training is not stability -> sigma in inference is our team struggle.
", "rawMarkdown": "@arabidopsisthalian we are also struggling to find the magic. at present -> training is not stability -> sigma in inference is our team struggle.", "votes": 3 } ], "comments": [ { "id": 2990550, "author_name": "Andrew Matteson", "author_url": "", "post_date": "2024-09-16T14:06:01.330000", "content": "Without giving anything away, my approach is along the lines of:
\nThen, the killer, reimplement everything in that pipeline to take advantage of this problem's structure to avoid bottlenecks in scipy/numpy. Otherwise, the test set would take ~200 hours to solve for.
\nI think my method will top out at about 0.63 unless I have some sort of additional insight.
", "votes": 18, "replies": [ { "id": 2990694, "author_name": "Georgii Aparin", "author_url": "", "post_date": "2024-09-16T16:09:58.353000", "content": "Thanks, how much \"cleaner\" signal improves your score? Or what do you mean by \"cleaner\" signal? As for me, no denoising algorithm has done well.
", "votes": 6, "replies": [ { "id": 2991013, "author_name": "Andrew Matteson", "author_url": "", "post_date": "2024-09-16T23:28:26.477000", "content": "I've tried close to a dozen approaches so far. There are small gains from some of them. The reason I put \"cleaner\" in quotes is, in some sense, simply averaging all the wavelengths provides you with a cleaner signal.
", "votes": 6, "replies": [] } ] } ] }, { "id": 2990498, "author_name": "greySnow", "author_url": "", "post_date": "2024-09-16T12:50:02.527000", "content": "I'm half hoping no one will reveal any 'magic' so I can stay in top 5, half hoping that someone will reveal exactly the 'magic' that I'm missing to get to 0.65+ range 🤣
", "votes": 7, "replies": [] }, { "id": 2990478, "author_name": "DennisSakva", "author_url": "", "post_date": "2024-09-16T12:22:28.070000", "content": "I'm pretty sure it has to do with a good estimation of the uncertainty part.
", "votes": 3, "replies": [ { "id": 2991112, "author_name": "ChingYinNg", "author_url": "", "post_date": "2024-09-17T04:15:50.533000", "content": "Yep. A very rough estimate of sigma boosted my score by 0.03
", "votes": 3, "replies": [] } ] }, { "id": 2990398, "author_name": "SeshuRaju 🧘♂️", "author_url": "", "post_date": "2024-09-16T11:01:46.637000", "content": "@arabidopsisthalian we are also struggling to find the magic. at present -> training is not stability -> sigma in inference is our team struggle.
", "votes": 3, "replies": [] } ], "raw_markdown_by_id": { "2990550": "Without giving anything away, my approach is along the lines of:\n- Get a \"cleaner\" signal.\n- Use optimization to get a best fit for depth.\n- Get a theoretical estimate of sigma from your best fit.\n- Model the noise given some basic domain knowledge. Get a different estimate of sigma.\n- Find a consensus estimate of sigma from your two approaches.\n\nThen, the killer, reimplement everything in that pipeline to take advantage of this problem's structure to avoid bottlenecks in scipy/numpy. Otherwise, the test set would take ~200 hours to solve for.\n\nI think my method will top out at about 0.63 unless I have some sort of additional insight.", "2990378": "What is the secret to achieving the top 10 LB places?\n\n\n\nPerhaps there is some information already in this discussion.\nWho can review everything we know about competition?\nIs there any important post list?", "2990498": "I'm half hoping no one will reveal any 'magic' so I can stay in top 5, half hoping that someone will reveal exactly the 'magic' that I'm missing to get to 0.65+ range 🤣", "2990478": "I'm pretty sure it has to do with a good estimation of the uncertainty part.", "2990398": "@arabidopsisthalian we are also struggling to find the magic. at present -> training is not stability -> sigma in inference is our team struggle." } }