![\"enter](\"https://i.ibb.co/Lv5Yn1s/score.png\")
First of all xgboost results are almost robust, it generates the same range of results each run. In addition, it can be enhanced with well-selected features.
\n\nThe first xgboost model published by me scored 0.614 then by adding 2 extra features, it reached to 0.623 in version 3 here. Which shows the potential to increase the score with more well selected features or params tuning. \"2 features increased = 0.09\"\n
Following features importance by xgboost model:\n![\"features\"](\"https://i.imgur.com/SpcFb19.png\")
If you suggest more features to test please add as comments
", "messages": [ { "id": "429863", "postDate": "11/29/2018 13:27:53", "content": "First of all xgboost results are almost robust, it generates the same range of results each run. In addition, it can be enhanced with well-selected features.
\n\nThe first xgboost model published by me scored 0.614 then by adding 2 extra features, it reached to 0.623 in version 3 here. Which shows the potential to increase the score with more well selected features or params tuning. \"2 features increased = 0.09\"\n
Following features importance by xgboost model:\n
If you suggest more features to test please add as comments
", "rawMarkdown": "First of all xgboost results are almost robust, it generates the same range of results each run. In addition, it can be enhanced with well-selected features.\n\nThe first xgboost model published by me scored 0.614 then by adding 2 extra features, it reached to **0.623** in version 3 [here][1]. Which shows the potential to increase the score with more well selected features or params tuning. *\"2 features increased = 0.09\"*\n![enter image description here][2]\n\nFollowing features importance by xgboost model:\n![features][3]\n\nIf you suggest more features to test please add as comments\n\n [1]: https://www.kaggle.com/jaguar00/xgboost-baseline\n [2]: https://i.ibb.co/Lv5Yn1s/score.png\n [3]: https://i.imgur.com/SpcFb19.png", "votes": null }, { "id": "435672", "postDate": "12/08/2018 14:13:51", "content": "You could try counting the number of some swear words in each sentence. Also, people tend to say 'you' before swearing or mention someone's race. Also, you might get a boost from checking if a question has a link, email, date or time.\nHope this helps!
", "rawMarkdown": "You could try counting the number of some swear words in each sentence. Also, people tend to say 'you' before swearing or mention someone's race. Also, you might get a boost from checking if a question has a link, email, date or time.\nHope this helps!", "votes": null }, { "id": "435753", "postDate": "12/08/2018 17:30:34", "content": "Nice ideas, going to test it and update you with results if any.
", "rawMarkdown": "Nice ideas, going to test it and update you with results if any.", "votes": null }, { "id": "444051", "postDate": "12/23/2018 02:43:11", "content": "Can such hand feature improve the NN score in your model?
", "rawMarkdown": "Can such hand feature improve the NN score in your model?", "votes": null }, { "id": "445244", "postDate": "12/26/2018 03:41:39", "content": "Not all the features maybe one or two, the 3rd place solution for similar competition \"Toxic\" used two features: https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/discussion/52644
", "rawMarkdown": "Not all the features maybe one or two, the 3rd place solution for similar competition \"Toxic\" used two features: https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/discussion/52644", "votes": null }, { "id": "457118", "postDate": "01/17/2019 01:13:39", "content": "Hi, did you get performance improvement after adding extra features? For me, I was unable to get extra feature working....
", "rawMarkdown": "Hi, did you get performance improvement after adding extra features? For me, I was unable to get extra feature working....", "votes": null }, { "id": "457167", "postDate": "01/17/2019 03:27:35", "content": "I tried other features such as the number of verbs and nouns however did not help to improve. Even tried it with RNN as an extra input and no enhancement.
", "rawMarkdown": "I tried other features such as the number of verbs and nouns however did not help to improve. Even tried it with RNN as an extra input and no enhancement.", "votes": null }, { "id": "458739", "postDate": "01/20/2019 12:03:41", "content": "Hi, I'm new kaggler. I use 300dims word embedding as input of first layer, and how to use such these statistics features in rnn? For example, when i get the number of verbs, should I take it after the word embedding and make it become 301 dims? Thank u so much!
", "rawMarkdown": "Hi, I'm new kaggler. I use 300dims word embedding as input of first layer, and how to use such these statistics features in rnn? For example, when i get the number of verbs, should I take it after the word embedding and make it become 301 dims? Thank u so much!", "votes": null }, { "id": "459580", "postDate": "01/22/2019 03:10:18", "content": "Have you tried POS, NER, syllables per word, readability indexes, sentiments, bad word list features?
", "rawMarkdown": "Have you tried POS, NER, syllables per word, readability indexes, sentiments, bad word list features?", "votes": null }, { "id": "459647", "postDate": "01/22/2019 06:04:31", "content": "Yes, not a much enhancement ... with RNN it showed a little enhancement although my best RNN model till now does not use it as input. \nCapital rate and unique words have a good impact on both XGboost and RNN based on mine testing.
", "rawMarkdown": "Yes, not a much enhancement ... with RNN it showed a little enhancement although my best RNN model till now does not use it as input. \nCapital rate and unique words have a good impact on both XGboost and RNN based on mine testing.", "votes": null }, { "id": "459649", "postDate": "01/22/2019 06:07:18", "content": "Thanks for your comment, following public kernel use multi-input for RNN (https://www.kaggle.com/jannen/reaching-0-7-fork-from-bilstm-attention-kfold) the code not much organized you have to read it carefully to avoid lines that consume time with no reason in this kernel.
", "rawMarkdown": "Thanks for your comment, following public kernel use multi-input for RNN (https://www.kaggle.com/jannen/reaching-0-7-fork-from-bilstm-attention-kfold) the code not much organized you have to read it carefully to avoid lines that consume time with no reason in this kernel.", "votes": null }, { "id": "461704", "postDate": "01/26/2019 19:34:09", "content": "I added some of these features my Local CV has improved while on LB it is performing worse.
", "rawMarkdown": "I added some of these features my Local CV has improved while on LB it is performing worse.", "votes": null } ], "comments": [ { "id": 435672, "author_name": "craq21", "author_url": "", "post_date": "12/08/2018 14:13:51", "content": "You could try counting the number of some swear words in each sentence. Also, people tend to say 'you' before swearing or mention someone's race. Also, you might get a boost from checking if a question has a link, email, date or time.\nHope this helps!
", "votes": null, "replies": [ { "id": 435753, "author_name": "jaguar00", "author_url": "", "post_date": "12/08/2018 17:30:34", "content": "Nice ideas, going to test it and update you with results if any.
", "votes": null, "replies": [] } ] }, { "id": 444051, "author_name": "codebb", "author_url": "", "post_date": "12/23/2018 02:43:11", "content": "Can such hand feature improve the NN score in your model?
", "votes": null, "replies": [ { "id": 445244, "author_name": "jaguar00", "author_url": "", "post_date": "12/26/2018 03:41:39", "content": "Not all the features maybe one or two, the 3rd place solution for similar competition \"Toxic\" used two features: https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/discussion/52644
", "votes": null, "replies": [] } ] }, { "id": 457118, "author_name": "zjucor", "author_url": "", "post_date": "01/17/2019 01:13:39", "content": "Hi, did you get performance improvement after adding extra features? For me, I was unable to get extra feature working....
", "votes": null, "replies": [ { "id": 457167, "author_name": "jaguar00", "author_url": "", "post_date": "01/17/2019 03:27:35", "content": "I tried other features such as the number of verbs and nouns however did not help to improve. Even tried it with RNN as an extra input and no enhancement.
", "votes": null, "replies": [] }, { "id": 461704, "author_name": "mlwhiz", "author_url": "", "post_date": "01/26/2019 19:34:09", "content": "I added some of these features my Local CV has improved while on LB it is performing worse.
", "votes": null, "replies": [] } ] }, { "id": 458739, "author_name": "karwik", "author_url": "", "post_date": "01/20/2019 12:03:41", "content": "Hi, I'm new kaggler. I use 300dims word embedding as input of first layer, and how to use such these statistics features in rnn? For example, when i get the number of verbs, should I take it after the word embedding and make it become 301 dims? Thank u so much!
", "votes": null, "replies": [ { "id": 459649, "author_name": "jaguar00", "author_url": "", "post_date": "01/22/2019 06:07:18", "content": "Thanks for your comment, following public kernel use multi-input for RNN (https://www.kaggle.com/jannen/reaching-0-7-fork-from-bilstm-attention-kfold) the code not much organized you have to read it carefully to avoid lines that consume time with no reason in this kernel.
", "votes": null, "replies": [] } ] }, { "id": 459580, "author_name": "kwigan", "author_url": "", "post_date": "01/22/2019 03:10:18", "content": "Have you tried POS, NER, syllables per word, readability indexes, sentiments, bad word list features?
", "votes": null, "replies": [ { "id": 459647, "author_name": "jaguar00", "author_url": "", "post_date": "01/22/2019 06:04:31", "content": "Yes, not a much enhancement ... with RNN it showed a little enhancement although my best RNN model till now does not use it as input. \nCapital rate and unique words have a good impact on both XGboost and RNN based on mine testing.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "429863": "First of all xgboost results are almost robust, it generates the same range of results each run. In addition, it can be enhanced with well-selected features.\n\nThe first xgboost model published by me scored 0.614 then by adding 2 extra features, it reached to **0.623** in version 3 [here][1]. Which shows the potential to increase the score with more well selected features or params tuning. *\"2 features increased = 0.09\"*\n![enter image description here][2]\n\nFollowing features importance by xgboost model:\n![features][3]\n\nIf you suggest more features to test please add as comments\n\n [1]: https://www.kaggle.com/jaguar00/xgboost-baseline\n [2]: https://i.ibb.co/Lv5Yn1s/score.png\n [3]: https://i.imgur.com/SpcFb19.png", "435672": "You could try counting the number of some swear words in each sentence. Also, people tend to say 'you' before swearing or mention someone's race. Also, you might get a boost from checking if a question has a link, email, date or time.\nHope this helps!", "435753": "Nice ideas, going to test it and update you with results if any.", "444051": "Can such hand feature improve the NN score in your model?", "445244": "Not all the features maybe one or two, the 3rd place solution for similar competition \"Toxic\" used two features: https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/discussion/52644", "457118": "Hi, did you get performance improvement after adding extra features? For me, I was unable to get extra feature working....", "457167": "I tried other features such as the number of verbs and nouns however did not help to improve. Even tried it with RNN as an extra input and no enhancement.", "458739": "Hi, I'm new kaggler. I use 300dims word embedding as input of first layer, and how to use such these statistics features in rnn? For example, when i get the number of verbs, should I take it after the word embedding and make it become 301 dims? Thank u so much!", "459580": "Have you tried POS, NER, syllables per word, readability indexes, sentiments, bad word list features?", "459647": "Yes, not a much enhancement ... with RNN it showed a little enhancement although my best RNN model till now does not use it as input. \nCapital rate and unique words have a good impact on both XGboost and RNN based on mine testing.", "459649": "Thanks for your comment, following public kernel use multi-input for RNN (https://www.kaggle.com/jannen/reaching-0-7-fork-from-bilstm-attention-kfold) the code not much organized you have to read it carefully to avoid lines that consume time with no reason in this kernel.", "461704": "I added some of these features my Local CV has improved while on LB it is performing worse." }, "source": "meta" }