I am trying to use additional data to improve the prediction and it seems to be very helpful. This data is presented only for the train dataset. Maybe I use not common approach (in ML competitions) but what is the common approach to use this type of data? I'm a novice in competitions, so I'm very interested in the information!
", "messages": [ { "id": "574947", "postDate": "07/14/2019 18:45:45", "content": "I am trying to use additional data to improve the prediction and it seems to be very helpful. This data is presented only for the train dataset. Maybe I use not common approach (in ML competitions) but what is the common approach to use this type of data? I'm a novice in competitions, so I'm very interested in the information!
", "rawMarkdown": "I am trying to use additional data to improve the prediction and it seems to be very helpful. This data is presented only for the train dataset. Maybe I use not common approach (in ML competitions) but what is the common approach to use this type of data? I'm a novice in competitions, so I'm very interested in the information!", "votes": null }, { "id": "575001", "postDate": "07/14/2019 20:35:01", "content": "Check out \"multitask-learning\". Can improve generalization, if you have problems with overfit... Might also be other benefits I am not aware of.
\n\nhttps://www.youtube.com/watch?v=UdXfsAr4Gjw
", "rawMarkdown": "Check out \"multitask-learning\". Can improve generalization, if you have problems with overfit... Might also be other benefits I am not aware of.\n\nhttps://www.youtube.com/watch?v=UdXfsAr4Gjw", "votes": null }, { "id": "575163", "postDate": "07/15/2019 06:01:52", "content": "As I understood I should add all this features as targets, so there will be 5 targets and learn a nn. Isn’t it? Thank you, going to try!
", "rawMarkdown": "As I understood I should add all this features as targets, so there will be 5 targets and learn a nn. Isn’t it? Thank you, going to try!", "votes": null }, { "id": "575178", "postDate": "07/15/2019 06:23:42", "content": "yes. Will force ther earlier layers to learn features that work for all targets.
\n\nPersonally, I haven't done this myself, since there is enough data for my NN models. But if I experience overfitting, I will do it.
", "rawMarkdown": "yes. Will force ther earlier layers to learn features that work for all targets. \n\nPersonally, I haven't done this myself, since there is enough data for my NN models. But if I experience overfitting, I will do it.", "votes": null }, { "id": "575221", "postDate": "07/15/2019 07:22:28", "content": "Which performance you can reach with nn?
", "rawMarkdown": "Which performance you can reach with nn?", "votes": null }, { "id": "575246", "postDate": "07/15/2019 07:58:28", "content": "dont know. im currently at -1.8.
", "rawMarkdown": "dont know. im currently at -1.8.", "votes": null }, { "id": "575392", "postDate": "07/15/2019 11:24:52", "content": "With qm9 data or without?
", "rawMarkdown": "With qm9 data or without?", "votes": null } ], "comments": [ { "id": 575001, "author_name": "makahana", "author_url": "", "post_date": "07/14/2019 20:35:01", "content": "Check out \"multitask-learning\". Can improve generalization, if you have problems with overfit... Might also be other benefits I am not aware of.
\n\nhttps://www.youtube.com/watch?v=UdXfsAr4Gjw
", "votes": null, "replies": [ { "id": 575163, "author_name": "kirill702b", "author_url": "", "post_date": "07/15/2019 06:01:52", "content": "As I understood I should add all this features as targets, so there will be 5 targets and learn a nn. Isn’t it? Thank you, going to try!
", "votes": null, "replies": [] }, { "id": 575178, "author_name": "makahana", "author_url": "", "post_date": "07/15/2019 06:23:42", "content": "yes. Will force ther earlier layers to learn features that work for all targets.
\n\nPersonally, I haven't done this myself, since there is enough data for my NN models. But if I experience overfitting, I will do it.
", "votes": null, "replies": [] }, { "id": 575221, "author_name": "kirill702b", "author_url": "", "post_date": "07/15/2019 07:22:28", "content": "Which performance you can reach with nn?
", "votes": null, "replies": [] }, { "id": 575246, "author_name": "makahana", "author_url": "", "post_date": "07/15/2019 07:58:28", "content": "dont know. im currently at -1.8.
", "votes": null, "replies": [] }, { "id": 575392, "author_name": "kirill702b", "author_url": "", "post_date": "07/15/2019 11:24:52", "content": "With qm9 data or without?
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "574947": "I am trying to use additional data to improve the prediction and it seems to be very helpful. This data is presented only for the train dataset. Maybe I use not common approach (in ML competitions) but what is the common approach to use this type of data? I'm a novice in competitions, so I'm very interested in the information!", "575001": "Check out \"multitask-learning\". Can improve generalization, if you have problems with overfit... Might also be other benefits I am not aware of.\n\nhttps://www.youtube.com/watch?v=UdXfsAr4Gjw", "575163": "As I understood I should add all this features as targets, so there will be 5 targets and learn a nn. Isn’t it? Thank you, going to try!", "575178": "yes. Will force ther earlier layers to learn features that work for all targets. \n\nPersonally, I haven't done this myself, since there is enough data for my NN models. But if I experience overfitting, I will do it.", "575221": "Which performance you can reach with nn?", "575246": "dont know. im currently at -1.8.", "575392": "With qm9 data or without?" }, "source": "meta" }