Hello! A couple of days ago I came up with two ideas: the use of LSTM and Category Encoders. LSTM made lots of sense, as patients have recurrent checkups. I wanted to if any others are using LSTM and how it compared to their other kernels. My model performed at -6.8652 with little to no adjustment. To see the Category Encoding thread for any responses to how well it worked, go to https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/173836
", "messages": [ { "id": "965902", "postDate": "08/11/2020 00:39:21", "content": "Hello! A couple of days ago I came up with two ideas: the use of LSTM and Category Encoders. LSTM made lots of sense, as patients have recurrent checkups. I wanted to if any others are using LSTM and how it compared to their other kernels. My model performed at -6.8652 with little to no adjustment. To see the Category Encoding thread for any responses to how well it worked, go to https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/173836
", "rawMarkdown": "Hello! A couple of days ago I came up with two ideas: the use of LSTM and Category Encoders. LSTM made lots of sense, as patients have recurrent checkups. I wanted to if any others are using LSTM and how it compared to their other kernels. My model performed at -6.8652 with little to no adjustment. To see the Category Encoding thread for any responses to how well it worked, go to https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/173836", "votes": null }, { "id": "967780", "postDate": "08/12/2020 13:57:02", "content": "Your link points to here. Is it correct?
", "rawMarkdown": "Your link points to here. Is it correct?", "votes": null }, { "id": "968031", "postDate": "08/12/2020 16:39:26", "content": "Yes, I linked the two subjects because I figured people might be interested in people's thoughts on these two ideas. But I still wanted the forum to be organized.
", "rawMarkdown": "Yes, I linked the two subjects because I figured people might be interested in people's thoughts on these two ideas. But I still wanted the forum to be organized.", "votes": null }, { "id": "968055", "postDate": "08/12/2020 17:06:13", "content": "CV -6.86 with the LSTM? I tried but couldn't get it to work because so much missing data. My set up was 146 (133--13) hidden states. Did you do the same or assume that each patient's measurements were sequential
", "rawMarkdown": "CV -6.86 with the LSTM? I tried but couldn't get it to work because so much missing data. My set up was 146 (133--13) hidden states. Did you do the same or assume that each patient's measurements were sequential", "votes": null }, { "id": "968056", "postDate": "08/12/2020 17:09:03", "content": "I agree that this would be a really elegant solution
", "rawMarkdown": "I agree that this would be a really elegant solution", "votes": null }, { "id": "968094", "postDate": "08/12/2020 17:37:36", "content": "I tried this in two ways: 1.Sequentially retraining in LSTM NN by feeding in subsequent samples, for example fit([patientsample1]) , fit([patientsample1,patientsample2]) …
\n2.loading in patient samples with a basic quantile regression schema but with an lstm layer
I think I'd be surprised if you got really good performance on the test data given it's invariant to the weeks the measurements are made and you've only got 1 measurement (the baseline) rather than a sequence.
\nI'd love to see a kernel with LSTM and a half-decent CV. I'd publish mine if I got below random 😂
", "rawMarkdown": "I think I'd be surprised if you got really good performance on the test data given it's invariant to the weeks the measurements are made and you've only got 1 measurement (the baseline) rather than a sequence. \n\nI'd love to see a kernel with LSTM and a half-decent CV. I'd publish mine if I got below random 😂", "votes": null }, { "id": "968935", "postDate": "08/13/2020 11:04:01", "content": "hi @eladwar can you release a LSTM starter notebook ?
", "rawMarkdown": "hi @eladwar can you release a LSTM starter notebook ?", "votes": null }, { "id": "969484", "postDate": "08/13/2020 17:59:30", "content": "I'll get on it, but I don't post notebooks too often. It may take me a while
", "rawMarkdown": "I'll get on it, but I don't post notebooks too often. It may take me a while", "votes": null }, { "id": "969663", "postDate": "08/13/2020 20:46:56", "content": "I'm about to post another LSTM attempt but for the dicom images, it will be a bit messy but hopefully you'll like it.
", "rawMarkdown": "I'm about to post another LSTM attempt but for the dicom images, it will be a bit messy but hopefully you'll like it.", "votes": null }, { "id": "1005709", "postDate": "09/10/2020 17:16:05", "content": "The conditional RNN notebook is at https://www.kaggle.com/eladwar/conditional-rnn?scriptVersionId=42399034 . Not as high a score as my lstm, but it's certainly more interesting.
", "rawMarkdown": "The conditional RNN notebook is at https://www.kaggle.com/eladwar/conditional-rnn?scriptVersionId=42399034 . Not as high a score as my lstm, but it's certainly more interesting.", "votes": null } ], "comments": [ { "id": 967780, "author_name": "atamazian", "author_url": "", "post_date": "08/12/2020 13:57:02", "content": "Your link points to here. Is it correct?
", "votes": null, "replies": [ { "id": 968031, "author_name": "eladwar", "author_url": "", "post_date": "08/12/2020 16:39:26", "content": "Yes, I linked the two subjects because I figured people might be interested in people's thoughts on these two ideas. But I still wanted the forum to be organized.
", "votes": null, "replies": [] }, { "id": 968055, "author_name": "jameschapman19", "author_url": "", "post_date": "08/12/2020 17:06:13", "content": "CV -6.86 with the LSTM? I tried but couldn't get it to work because so much missing data. My set up was 146 (133--13) hidden states. Did you do the same or assume that each patient's measurements were sequential
", "votes": null, "replies": [] }, { "id": 968056, "author_name": "jameschapman19", "author_url": "", "post_date": "08/12/2020 17:09:03", "content": "I agree that this would be a really elegant solution
", "votes": null, "replies": [] }, { "id": 968094, "author_name": "eladwar", "author_url": "", "post_date": "08/12/2020 17:37:36", "content": "I tried this in two ways: 1.Sequentially retraining in LSTM NN by feeding in subsequent samples, for example fit([patientsample1]) , fit([patientsample1,patientsample2]) …
\n2.loading in patient samples with a basic quantile regression schema but with an lstm layer
I think I'd be surprised if you got really good performance on the test data given it's invariant to the weeks the measurements are made and you've only got 1 measurement (the baseline) rather than a sequence.
\nI'd love to see a kernel with LSTM and a half-decent CV. I'd publish mine if I got below random 😂
", "votes": null, "replies": [] } ] }, { "id": 968935, "author_name": "ulrich07", "author_url": "", "post_date": "08/13/2020 11:04:01", "content": "hi @eladwar can you release a LSTM starter notebook ?
", "votes": null, "replies": [ { "id": 969484, "author_name": "eladwar", "author_url": "", "post_date": "08/13/2020 17:59:30", "content": "I'll get on it, but I don't post notebooks too often. It may take me a while
", "votes": null, "replies": [] }, { "id": 969663, "author_name": "eladwar", "author_url": "", "post_date": "08/13/2020 20:46:56", "content": "I'm about to post another LSTM attempt but for the dicom images, it will be a bit messy but hopefully you'll like it.
", "votes": null, "replies": [] }, { "id": 1005709, "author_name": "eladwar", "author_url": "", "post_date": "09/10/2020 17:16:05", "content": "The conditional RNN notebook is at https://www.kaggle.com/eladwar/conditional-rnn?scriptVersionId=42399034 . Not as high a score as my lstm, but it's certainly more interesting.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "965902": "Hello! A couple of days ago I came up with two ideas: the use of LSTM and Category Encoders. LSTM made lots of sense, as patients have recurrent checkups. I wanted to if any others are using LSTM and how it compared to their other kernels. My model performed at -6.8652 with little to no adjustment. To see the Category Encoding thread for any responses to how well it worked, go to https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/173836", "967780": "Your link points to here. Is it correct?", "968031": "Yes, I linked the two subjects because I figured people might be interested in people's thoughts on these two ideas. But I still wanted the forum to be organized.", "968055": "CV -6.86 with the LSTM? I tried but couldn't get it to work because so much missing data. My set up was 146 (133--13) hidden states. Did you do the same or assume that each patient's measurements were sequential", "968056": "I agree that this would be a really elegant solution", "968094": "I tried this in two ways: 1.Sequentially retraining in LSTM NN by feeding in subsequent samples, for example fit([patient_sample_1]) , fit([patient_sample_1,patient_sample_2]) ...\n2.loading in patient samples with a basic quantile regression schema but with an lstm layer", "968273": "I think I'd be surprised if you got really good performance on the test data given it's invariant to the weeks the measurements are made and you've only got 1 measurement (the baseline) rather than a sequence. \n\nI'd love to see a kernel with LSTM and a half-decent CV. I'd publish mine if I got below random 😂", "968935": "hi @eladwar can you release a LSTM starter notebook ?", "969484": "I'll get on it, but I don't post notebooks too often. It may take me a while", "969663": "I'm about to post another LSTM attempt but for the dicom images, it will be a bit messy but hopefully you'll like it.", "1005709": "The conditional RNN notebook is at https://www.kaggle.com/eladwar/conditional-rnn?scriptVersionId=42399034 . Not as high a score as my lstm, but it's certainly more interesting." }, "source": "meta" }