I was trying to run simple model architectures on this dataset, i cleaned text using few basic regexes, made an embedding layer and then tried below architectures,\n1) embedding + avgpool + dense + dense sigmoid\n2)embedding + lstm + dense sigmoid\nfor both types my model keeps on predicting 0's for each input. Is it becuase of class imbalance or i am making mistake in model i am unsure.Please give me few pointers to begin.
", "messages": [ { "id": "437785", "postDate": "12/12/2018 13:43:04", "content": "I was trying to run simple model architectures on this dataset, i cleaned text using few basic regexes, made an embedding layer and then tried below architectures,\n1) embedding + avgpool + dense + dense sigmoid\n2)embedding + lstm + dense sigmoid\nfor both types my model keeps on predicting 0's for each input. Is it becuase of class imbalance or i am making mistake in model i am unsure.Please give me few pointers to begin.
", "rawMarkdown": "I was trying to run simple model architectures on this dataset, i cleaned text using few basic regexes, made an embedding layer and then tried below architectures,\n1) embedding + avgpool + dense + dense sigmoid\n2)embedding + lstm + dense sigmoid\nfor both types my model keeps on predicting 0's for each input. Is it becuase of class imbalance or i am making mistake in model i am unsure.Please give me few pointers to begin.", "votes": null }, { "id": "437842", "postDate": "12/12/2018 15:51:33", "content": "I think a little more information is needed before knowing for sure. First, i would make sure you have all the needed base layers: input -> embedding -> pooling or flatten -> dense (maybe use relu) -> dense (sigmoid). Secondly, It could be coming from overfitting if you're using too small of a batch size or too big of a epoch size. Lastly, as shown in my kernel (https://www.kaggle.com/mgancita/logit-regression-detailed-building-of-lstm-model?scriptVersionId=8098690 ) only about 6.2% of the questions are insincere so this imbalance may be causing your model to be overfit to always predict 0.
", "rawMarkdown": "I think a little more information is needed before knowing for sure. First, i would make sure you have all the needed base layers: input -> embedding -> pooling or flatten -> dense (maybe use relu) -> dense (sigmoid). Secondly, It could be coming from overfitting if you're using too small of a batch size or too big of a epoch size. Lastly, as shown in my kernel (https://www.kaggle.com/mgancita/logit-regression-detailed-building-of-lstm-model?scriptVersionId=8098690 ) only about 6.2% of the questions are insincere so this imbalance may be causing your model to be overfit to always predict 0.", "votes": null }, { "id": "438135", "postDate": "12/13/2018 06:44:06", "content": "i am exactly using input -> embedding -> pooling or flatten -> dense (maybe use relu) -> dense (sigmoid) type of structure.I tired varying batch size(256-512-1024) and epochs(1-10). \nAbout overfitting there are other kernels out there who also use simple lstm models and yet they get atleast 0.60 score not sure why i would get all 0's
", "rawMarkdown": "i am exactly using input -> embedding -> pooling or flatten -> dense (maybe use relu) -> dense (sigmoid) type of structure.I tired varying batch size(256-512-1024) and epochs(1-10). \nAbout overfitting there are other kernels out there who also use simple lstm models and yet they get atleast 0.60 score not sure why i would get all 0's", "votes": null }, { "id": "438664", "postDate": "12/14/2018 02:07:16", "content": "I have not gone through your code. But I believe a big part of this is the imbalance between labels 0 and 1. This is known as the imbalanced classification problem. Because one label is in a much larger number than the others, the model has failed to learn and simply predict the majority class label.
\n\nWhat you can do is to resample the data.
", "rawMarkdown": "I have not gone through your code. But I believe a big part of this is the imbalance between labels 0 and 1. This is known as the imbalanced classification problem. Because one label is in a much larger number than the others, the model has failed to learn and simply predict the majority class label.\n\nWhat you can do is to resample the data.", "votes": null }, { "id": "439292", "postDate": "12/15/2018 05:13:58", "content": "i get the imbalanced classification.But there are other kernels out here , who use similar simple arcitecture and yet get somewhere in region of 60%,\ntake this kernel for example \nhttps://www.kaggle.com/nitinaggarwal008/basic-model\ni examined and found same imbalance in train label sets\n 0 1127298\n1 74334
", "rawMarkdown": "i get the imbalanced classification.But there are other kernels out here , who use similar simple arcitecture and yet get somewhere in region of 60%,\ntake this kernel for example \nhttps://www.kaggle.com/nitinaggarwal008/basic-model\ni examined and found same imbalance in train label sets\n 0 1127298\n1 74334", "votes": null }, { "id": "440054", "postDate": "12/17/2018 00:58:55", "content": "Perhaps you can post the link to your kernel for us to see?
", "rawMarkdown": "Perhaps you can post the link to your kernel for us to see?", "votes": null }, { "id": "443693", "postDate": "12/22/2018 07:05:44", "content": "Got it resolved.I was not properly reading glove vectors in my Embedding layer.My embedding layer was always np.zeros array and hence the all 0 predictions.Feeling very stupid now.
", "rawMarkdown": "Got it resolved.I was not properly reading glove vectors in my Embedding layer.My embedding layer was always np.zeros array and hence the all 0 predictions.Feeling very stupid now.", "votes": null } ], "comments": [ { "id": 437842, "author_name": "mgancita", "author_url": "", "post_date": "12/12/2018 15:51:33", "content": "I think a little more information is needed before knowing for sure. First, i would make sure you have all the needed base layers: input -> embedding -> pooling or flatten -> dense (maybe use relu) -> dense (sigmoid). Secondly, It could be coming from overfitting if you're using too small of a batch size or too big of a epoch size. Lastly, as shown in my kernel (https://www.kaggle.com/mgancita/logit-regression-detailed-building-of-lstm-model?scriptVersionId=8098690 ) only about 6.2% of the questions are insincere so this imbalance may be causing your model to be overfit to always predict 0.
", "votes": null, "replies": [ { "id": 438135, "author_name": "chansbest93", "author_url": "", "post_date": "12/13/2018 06:44:06", "content": "i am exactly using input -> embedding -> pooling or flatten -> dense (maybe use relu) -> dense (sigmoid) type of structure.I tired varying batch size(256-512-1024) and epochs(1-10). \nAbout overfitting there are other kernels out there who also use simple lstm models and yet they get atleast 0.60 score not sure why i would get all 0's
", "votes": null, "replies": [] } ] }, { "id": 438664, "author_name": "gxkok21", "author_url": "", "post_date": "12/14/2018 02:07:16", "content": "I have not gone through your code. But I believe a big part of this is the imbalance between labels 0 and 1. This is known as the imbalanced classification problem. Because one label is in a much larger number than the others, the model has failed to learn and simply predict the majority class label.
\n\nWhat you can do is to resample the data.
", "votes": null, "replies": [ { "id": 439292, "author_name": "chansbest93", "author_url": "", "post_date": "12/15/2018 05:13:58", "content": "i get the imbalanced classification.But there are other kernels out here , who use similar simple arcitecture and yet get somewhere in region of 60%,\ntake this kernel for example \nhttps://www.kaggle.com/nitinaggarwal008/basic-model\ni examined and found same imbalance in train label sets\n 0 1127298\n1 74334
", "votes": null, "replies": [] }, { "id": 440054, "author_name": "gxkok21", "author_url": "", "post_date": "12/17/2018 00:58:55", "content": "Perhaps you can post the link to your kernel for us to see?
", "votes": null, "replies": [] }, { "id": 443693, "author_name": "chansbest93", "author_url": "", "post_date": "12/22/2018 07:05:44", "content": "Got it resolved.I was not properly reading glove vectors in my Embedding layer.My embedding layer was always np.zeros array and hence the all 0 predictions.Feeling very stupid now.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "437785": "I was trying to run simple model architectures on this dataset, i cleaned text using few basic regexes, made an embedding layer and then tried below architectures,\n1) embedding + avgpool + dense + dense sigmoid\n2)embedding + lstm + dense sigmoid\nfor both types my model keeps on predicting 0's for each input. Is it becuase of class imbalance or i am making mistake in model i am unsure.Please give me few pointers to begin.", "437842": "I think a little more information is needed before knowing for sure. First, i would make sure you have all the needed base layers: input -> embedding -> pooling or flatten -> dense (maybe use relu) -> dense (sigmoid). Secondly, It could be coming from overfitting if you're using too small of a batch size or too big of a epoch size. Lastly, as shown in my kernel (https://www.kaggle.com/mgancita/logit-regression-detailed-building-of-lstm-model?scriptVersionId=8098690 ) only about 6.2% of the questions are insincere so this imbalance may be causing your model to be overfit to always predict 0.", "438135": "i am exactly using input -> embedding -> pooling or flatten -> dense (maybe use relu) -> dense (sigmoid) type of structure.I tired varying batch size(256-512-1024) and epochs(1-10). \nAbout overfitting there are other kernels out there who also use simple lstm models and yet they get atleast 0.60 score not sure why i would get all 0's", "438664": "I have not gone through your code. But I believe a big part of this is the imbalance between labels 0 and 1. This is known as the imbalanced classification problem. Because one label is in a much larger number than the others, the model has failed to learn and simply predict the majority class label.\n\nWhat you can do is to resample the data.", "439292": "i get the imbalanced classification.But there are other kernels out here , who use similar simple arcitecture and yet get somewhere in region of 60%,\ntake this kernel for example \nhttps://www.kaggle.com/nitinaggarwal008/basic-model\ni examined and found same imbalance in train label sets\n 0 1127298\n1 74334", "440054": "Perhaps you can post the link to your kernel for us to see?", "443693": "Got it resolved.I was not properly reading glove vectors in my Embedding layer.My embedding layer was always np.zeros array and hence the all 0 predictions.Feeling very stupid now." }, "source": "meta" }