I am trying to come up with a model for predicting earth quakes. I have tried both xgboost and deep neural network based approach, I find that deep learning methods are consistently underperforming the other algorithms. Can anyone shed light on the above?
", "messages": [ { "id": 494354, "postDate": "2019-03-19T18:34:24.140Z", "content": "I used random forest to get my highest score. I'm currently experimenting with xgb though.
", "rawMarkdown": "I used random forest to get my highest score. I'm currently experimenting with xgb though.", "votes": 11, "replies": [ { "id": 521523, "postDate": "2019-04-23T02:00:44.723Z", "content": "Using only random forest to achieve such a high score. Your feature engineering must be excellent.
", "rawMarkdown": "Using only random forest to achieve such a high score. Your feature engineering must be excellent." }, { "id": 521551, "postDate": "2019-04-23T03:16:12.183Z", "content": "Random Forest is better for noisy datasets
", "rawMarkdown": "Random Forest is better for noisy datasets", "votes": 3 } ] }, { "id": 494086, "postDate": "2019-03-19T12:56:14.097Z", "content": "In order to extract maximum from DL you need to preserve the original resolution of the raw and it's not trivial from the memory point of view. After feature extraction there is no any value for DL, so in this case (like all do) xgb wins.
", "rawMarkdown": "In order to extract maximum from DL you need to preserve the original resolution of the raw and it's not trivial from the memory point of view. After feature extraction there is no any value for DL, so in this case (like all do) xgb wins.", "votes": 8, "replies": [ { "id": 521553, "postDate": "2019-04-23T03:19:01.807Z", "content": "You can use attention mechanisms, or use Temporal Convolutional Networks like me. However I cannot crack 2.08 CV on 5-fold. Lots of dropout is required too. I have some new ideas for features that hopefully will give me strength
", "rawMarkdown": "You can use attention mechanisms, or use Temporal Convolutional Networks like me. However I cannot crack 2.08 CV on 5-fold. Lots of dropout is required too. I have some new ideas for features that hopefully will give me strength", "votes": 1 } ] }, { "id": 493849, "postDate": "2019-03-19T06:56:31.277Z", "content": "I am trying to come up with a model for predicting earth quakes. I have tried both xgboost and deep neural network based approach, I find that deep learning methods are consistently underperforming the other algorithms. Can anyone shed light on the above?
", "rawMarkdown": "I am trying to come up with a model for predicting earth quakes. I have tried both xgboost and deep neural network based approach, I find that deep learning methods are consistently underperforming the other algorithms. Can anyone shed light on the above?", "votes": 5 }, { "id": 521687, "postDate": "2019-04-23T09:11:40.987Z", "content": "I bet top teams mostly are using nn models.
", "rawMarkdown": "I bet top teams mostly are using nn models.", "votes": 1, "replies": [ { "id": 521705, "postDate": "2019-04-23T09:57:31.977Z", "content": "NN within an ensemble IMHO.
", "rawMarkdown": "NN within an ensemble IMHO.", "votes": 1 }, { "id": 524562, "postDate": "2019-04-29T05:11:03.567Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 497884, "postDate": "2019-03-24T03:01:23.010Z", "content": "i only tried DL without much feature engineering, it performs bad so far.
", "rawMarkdown": "i only tried DL without much feature engineering, it performs bad so far. ", "votes": 1 }, { "id": 521436, "postDate": "2019-04-22T23:23:23.080Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 494354, "author_name": "Machinehead", "author_url": "", "post_date": "2019-03-19T18:34:24.140000", "content": "I used random forest to get my highest score. I'm currently experimenting with xgb though.
", "votes": 11, "replies": [ { "id": 521523, "author_name": "FP", "author_url": "", "post_date": "2019-04-23T02:00:44.723000", "content": "Using only random forest to achieve such a high score. Your feature engineering must be excellent.
", "votes": 0, "replies": [] }, { "id": 521551, "author_name": "CoreyJamesLevinson", "author_url": "", "post_date": "2019-04-23T03:16:12.183000", "content": "Random Forest is better for noisy datasets
", "votes": 3, "replies": [] } ] }, { "id": 494086, "author_name": "", "author_url": "", "post_date": "2019-03-19T12:56:14.097000", "content": "In order to extract maximum from DL you need to preserve the original resolution of the raw and it's not trivial from the memory point of view. After feature extraction there is no any value for DL, so in this case (like all do) xgb wins.
", "votes": 8, "replies": [ { "id": 521553, "author_name": "CoreyJamesLevinson", "author_url": "", "post_date": "2019-04-23T03:19:01.807000", "content": "You can use attention mechanisms, or use Temporal Convolutional Networks like me. However I cannot crack 2.08 CV on 5-fold. Lots of dropout is required too. I have some new ideas for features that hopefully will give me strength
", "votes": 1, "replies": [] } ] }, { "id": 521687, "author_name": "Marcus Lin", "author_url": "", "post_date": "2019-04-23T09:11:40.987000", "content": "I bet top teams mostly are using nn models.
", "votes": 1, "replies": [ { "id": 521705, "author_name": "CPMP", "author_url": "", "post_date": "2019-04-23T09:57:31.977000", "content": "NN within an ensemble IMHO.
", "votes": 1, "replies": [] }, { "id": 524562, "author_name": "", "author_url": "", "post_date": "2019-04-29T05:11:03.567000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 497884, "author_name": "Z. Liu", "author_url": "", "post_date": "2019-03-24T03:01:23.010000", "content": "i only tried DL without much feature engineering, it performs bad so far.
", "votes": 1, "replies": [] }, { "id": 521436, "author_name": "", "author_url": "", "post_date": "2019-04-22T23:23:23.080000", "content": "", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "494354": "I used random forest to get my highest score. I'm currently experimenting with xgb though.", "494086": "In order to extract maximum from DL you need to preserve the original resolution of the raw and it's not trivial from the memory point of view. After feature extraction there is no any value for DL, so in this case (like all do) xgb wins.", "493849": "I am trying to come up with a model for predicting earth quakes. I have tried both xgboost and deep neural network based approach, I find that deep learning methods are consistently underperforming the other algorithms. Can anyone shed light on the above?", "521687": "I bet top teams mostly are using nn models.", "497884": "i only tried DL without much feature engineering, it performs bad so far. ", "521436": "" } }