it is my first competition, and our term focus on models.
\n\nnow, we have release our model in https://www.kaggle.com/xiaobai1123q/15th-place-solution
\n\nwe run our model again ( because the submitted kernel is a version ), and achieve a better result than leaderboard. I hope we can bring you some help.
\n\nin the text preprocessing stage, we don't have any personal work, all of which are public kernels.
\n\nour main job lies in the four models we ensemble. then, i will briefly explain.
\n\nthe first model is RCNN.\nthe second model is LSTM(128) + GRU(96) + maxpooling1D + dropout(0.1).\nthe third model is LSTM(128) + GRU(64) + Conv1D + maxpooling_concatenate.\nthe fourth model is LSTM(128) + GRU(64) + Conv1D + Attention.
\n\nwe used the word vector concatenated by glove and fasttext.\nwe set max_features = None and we set max_len = 57.
\n\nquestions, advises, suggestions are all welcome.
", "messages": [ { "id": "471306", "postDate": "02/14/2019 09:40:43", "content": "it is my first competition, and our term focus on models.
\n\nnow, we have release our model in https://www.kaggle.com/xiaobai1123q/15th-place-solution
\n\nwe run our model again ( because the submitted kernel is a version ), and achieve a better result than leaderboard. I hope we can bring you some help.
\n\nin the text preprocessing stage, we don't have any personal work, all of which are public kernels.
\n\nour main job lies in the four models we ensemble. then, i will briefly explain.
\n\nthe first model is RCNN.\nthe second model is LSTM(128) + GRU(96) + maxpooling1D + dropout(0.1).\nthe third model is LSTM(128) + GRU(64) + Conv1D + maxpooling_concatenate.\nthe fourth model is LSTM(128) + GRU(64) + Conv1D + Attention.
\n\nwe used the word vector concatenated by glove and fasttext.\nwe set max_features = None and we set max_len = 57.
\n\nquestions, advises, suggestions are all welcome.
", "rawMarkdown": "it is my first competition, and our term focus on models.\n\nnow, we have release our model in https://www.kaggle.com/xiaobai1123q/15th-place-solution \n\nwe run our model again ( because the submitted kernel is a version ), and achieve a better result than leaderboard. I hope we can bring you some help.\n\nin the text preprocessing stage, we don't have any personal work, all of which are public kernels.\n\nour main job lies in the four models we ensemble. then, i will briefly explain.\n\nthe first model is RCNN.\nthe second model is LSTM(128) + GRU(96) + maxpooling1D + dropout(0.1).\nthe third model is LSTM(128) + GRU(64) + Conv1D + maxpooling_concatenate.\nthe fourth model is LSTM(128) + GRU(64) + Conv1D + Attention.\n\nwe used the word vector concatenated by glove and fasttext.\nwe set max_features = None and we set max_len = 57.\n\nquestions, advises, suggestions are all welcome.", "votes": null }, { "id": "471319", "postDate": "02/14/2019 09:52:49", "content": "How did you chose the weights for averaging? Have you tried any other ensembling methods like multilayered perceptrons or LGBM?
", "rawMarkdown": "How did you chose the weights for averaging? Have you tried any other ensembling methods like multilayered perceptrons or LGBM?", "votes": null }, { "id": "471323", "postDate": "02/14/2019 09:56:09", "content": "Congratulations @bai. Simple solution, Keep up the good work. What was the CV score without RCNN? Did RCNN boost CV score?
", "rawMarkdown": "Congratulations @bai. Simple solution, Keep up the good work. What was the CV score without RCNN? Did RCNN boost CV score?", "votes": null }, { "id": "471325", "postDate": "02/14/2019 09:57:09", "content": "we do a lot of experiments on it, simple weights are the best choice. the weights are based on our observations and determined by humans.
", "rawMarkdown": "we do a lot of experiments on it, simple weights are the best choice. the weights are based on our observations and determined by humans.", "votes": null }, { "id": "471328", "postDate": "02/14/2019 09:59:53", "content": "thank you. if remove any of these models, the final performance will drop dramatically, not just RCNN.\nand we do simple splits, not CV. because of the time limit.
", "rawMarkdown": "thank you. if remove any of these models, the final performance will drop dramatically, not just RCNN.\nand we do simple splits, not CV. because of the time limit.", "votes": null }, { "id": "471360", "postDate": "02/14/2019 10:55:52", "content": "Congrats on finishing in gold range!
", "rawMarkdown": "Congrats on finishing in gold range!", "votes": null }, { "id": "471848", "postDate": "02/15/2019 01:33:43", "content": "thank you!
", "rawMarkdown": "thank you!", "votes": null } ], "comments": [ { "id": 471319, "author_name": "anasuna", "author_url": "", "post_date": "02/14/2019 09:52:49", "content": "How did you chose the weights for averaging? Have you tried any other ensembling methods like multilayered perceptrons or LGBM?
", "votes": null, "replies": [ { "id": 471325, "author_name": "xiaobai1123q", "author_url": "", "post_date": "02/14/2019 09:57:09", "content": "we do a lot of experiments on it, simple weights are the best choice. the weights are based on our observations and determined by humans.
", "votes": null, "replies": [] } ] }, { "id": 471323, "author_name": "karthik7395", "author_url": "", "post_date": "02/14/2019 09:56:09", "content": "Congratulations @bai. Simple solution, Keep up the good work. What was the CV score without RCNN? Did RCNN boost CV score?
", "votes": null, "replies": [ { "id": 471328, "author_name": "xiaobai1123q", "author_url": "", "post_date": "02/14/2019 09:59:53", "content": "thank you. if remove any of these models, the final performance will drop dramatically, not just RCNN.\nand we do simple splits, not CV. because of the time limit.
", "votes": null, "replies": [] } ] }, { "id": 471360, "author_name": "springmanndaniel", "author_url": "", "post_date": "02/14/2019 10:55:52", "content": "Congrats on finishing in gold range!
", "votes": null, "replies": [ { "id": 471848, "author_name": "xiaobai1123q", "author_url": "", "post_date": "02/15/2019 01:33:43", "content": "thank you!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "471306": "it is my first competition, and our term focus on models.\n\nnow, we have release our model in https://www.kaggle.com/xiaobai1123q/15th-place-solution \n\nwe run our model again ( because the submitted kernel is a version ), and achieve a better result than leaderboard. I hope we can bring you some help.\n\nin the text preprocessing stage, we don't have any personal work, all of which are public kernels.\n\nour main job lies in the four models we ensemble. then, i will briefly explain.\n\nthe first model is RCNN.\nthe second model is LSTM(128) + GRU(96) + maxpooling1D + dropout(0.1).\nthe third model is LSTM(128) + GRU(64) + Conv1D + maxpooling_concatenate.\nthe fourth model is LSTM(128) + GRU(64) + Conv1D + Attention.\n\nwe used the word vector concatenated by glove and fasttext.\nwe set max_features = None and we set max_len = 57.\n\nquestions, advises, suggestions are all welcome.", "471319": "How did you chose the weights for averaging? Have you tried any other ensembling methods like multilayered perceptrons or LGBM?", "471323": "Congratulations @bai. Simple solution, Keep up the good work. What was the CV score without RCNN? Did RCNN boost CV score?", "471325": "we do a lot of experiments on it, simple weights are the best choice. the weights are based on our observations and determined by humans.", "471328": "thank you. if remove any of these models, the final performance will drop dramatically, not just RCNN.\nand we do simple splits, not CV. because of the time limit.", "471360": "Congrats on finishing in gold range!", "471848": "thank you!" }, "source": "meta" }