When I first started this competition I had strong faith that I would be successful with a neural net with embeddings, but never got above .97 LB. I have since switched to lgbm and been successful, but I'm curious if anyone else has had some success with NNs?
", "messages": [ { "id": "318729", "postDate": "04/24/2018 10:46:29", "content": "When I first started this competition I had strong faith that I would be successful with a neural net with embeddings, but never got above .97 LB. I have since switched to lgbm and been successful, but I'm curious if anyone else has had some success with NNs?
", "rawMarkdown": "When I first started this competition I had strong faith that I would be successful with a neural net with embeddings, but never got above .97 LB. I have since switched to lgbm and been successful, but I'm curious if anyone else has had some success with NNs?", "votes": null }, { "id": "318950", "postDate": "04/24/2018 20:19:35", "content": "I'm trying to construct a NN in the light of https://www.kaggle.com/whitebird/mercari-price-3rd-0-3905-cv-at-pb-in-3300-s but I didn't got interesting result yet.
", "rawMarkdown": "I'm trying to construct a NN in the light of https://www.kaggle.com/whitebird/mercari-price-3rd-0-3905-cv-at-pb-in-3300-s but I didn't got interesting result yet.", "votes": null }, { "id": "319515", "postDate": "04/26/2018 07:56:46", "content": "I'm focusing on NN using embeddings and added a concept similar to attention based on generated features but so far my LB ROC AUC is not that great. That's probably because I'm stubborn and I don't want to create features that relay on data ahead in the future (e.g. data based on is_attributed).
\n\nMy local ROC AUC, based on my validation set is over 0.98. I guess my NN is over-fitting. It's strange though, because I've managed to get my network down to 197865 params and wouldn't have expected it to over-fit.
\n\nStrangely enough my NN gets to a platou after 1 epoch (~ 30 minutes on a GTX 1070 G1) and lowering the LR doesn't improve the results. This means I'm free to re-train and re-sample train data and validation data.
", "rawMarkdown": "I'm focusing on NN using embeddings and added a concept similar to attention based on generated features but so far my LB ROC AUC is not that great. That's probably because I'm stubborn and I don't want to create features that relay on data ahead in the future (e.g. data based on is_attributed). \n\nMy local ROC AUC, based on my validation set is over 0.98. I guess my NN is over-fitting. It's strange though, because I've managed to get my network down to 197865 params and wouldn't have expected it to over-fit.\n\nStrangely enough my NN gets to a platou after 1 epoch (~ 30 minutes on a GTX 1070 G1) and lowering the LR doesn't improve the results. This means I'm free to re-train and re-sample train data and validation data.", "votes": null }, { "id": "319530", "postDate": "04/26/2018 08:35:54", "content": "See https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/54929#319416
", "rawMarkdown": "See https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/54929#319416", "votes": null } ], "comments": [ { "id": 318950, "author_name": "marrvolo", "author_url": "", "post_date": "04/24/2018 20:19:35", "content": "I'm trying to construct a NN in the light of https://www.kaggle.com/whitebird/mercari-price-3rd-0-3905-cv-at-pb-in-3300-s but I didn't got interesting result yet.
", "votes": null, "replies": [] }, { "id": 319515, "author_name": "profetul", "author_url": "", "post_date": "04/26/2018 07:56:46", "content": "I'm focusing on NN using embeddings and added a concept similar to attention based on generated features but so far my LB ROC AUC is not that great. That's probably because I'm stubborn and I don't want to create features that relay on data ahead in the future (e.g. data based on is_attributed).
\n\nMy local ROC AUC, based on my validation set is over 0.98. I guess my NN is over-fitting. It's strange though, because I've managed to get my network down to 197865 params and wouldn't have expected it to over-fit.
\n\nStrangely enough my NN gets to a platou after 1 epoch (~ 30 minutes on a GTX 1070 G1) and lowering the LR doesn't improve the results. This means I'm free to re-train and re-sample train data and validation data.
", "votes": null, "replies": [] }, { "id": 319530, "author_name": "cpmpml", "author_url": "", "post_date": "04/26/2018 08:35:54", "content": "See https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/54929#319416
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "318729": "When I first started this competition I had strong faith that I would be successful with a neural net with embeddings, but never got above .97 LB. I have since switched to lgbm and been successful, but I'm curious if anyone else has had some success with NNs?", "318950": "I'm trying to construct a NN in the light of https://www.kaggle.com/whitebird/mercari-price-3rd-0-3905-cv-at-pb-in-3300-s but I didn't got interesting result yet.", "319515": "I'm focusing on NN using embeddings and added a concept similar to attention based on generated features but so far my LB ROC AUC is not that great. That's probably because I'm stubborn and I don't want to create features that relay on data ahead in the future (e.g. data based on is_attributed). \n\nMy local ROC AUC, based on my validation set is over 0.98. I guess my NN is over-fitting. It's strange though, because I've managed to get my network down to 197865 params and wouldn't have expected it to over-fit.\n\nStrangely enough my NN gets to a platou after 1 epoch (~ 30 minutes on a GTX 1070 G1) and lowering the LR doesn't improve the results. This means I'm free to re-train and re-sample train data and validation data.", "319530": "See https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/54929#319416" }, "source": "meta" }