Since there is no lable train/test_active.csv. Text information can be collected as a corpus. What can be done to the others? Any advice will thank a lot.
", "messages": [ { "id": 327605, "postDate": "2018-05-12T01:29:28.557Z", "content": "Since there is no lable train/test_active.csv. Text information can be collected as a corpus. What can be done to the others? Any advice will thank a lot.
", "rawMarkdown": "Since there is no lable train/test_active.csv. Text information can be collected as a corpus. What can be done to the others? Any advice will thank a lot.\n\n", "votes": 9 }, { "id": 327911, "postDate": "2018-05-12T21:51:52.403Z", "content": "Currently using No.1 in my model.
\n\nHaven't tried No.2, though there's a nicely upvoted kernel that does essentially this.
\n\nNo.3 failed for me. I basically tried to follow Michael Jahrer's approach (from Porto) with a DAE but instead of using sparse features, I used an embedding on the categorical variable. Other changes were I also included some statistical features in the input which were not recreated in the reconstruction. Lastly, I implemented a custom Dense layer, which transposed the initial embedding matrix, essentially tying the weights and halving the model complexity. Maybe I bit off more than I can chew, or maybe I don't know what I'm doing at all. Trained with sparse cross entropy and the model learned... something because the final loss was about 20-25% it's initial loss. I only kept the categorical embeddings, threw them into my model as untrainable and ran it again. Loss dropped... but didn't beat--nor even come close--to my baseline.
\n\nIf anyone else gives no.3 a whirl, please let me know how it goes. I tried training on train/train_a/test/test_a combined, for the available columns.
\n\nAnd per @aquatic above, I will eventually try PL'ing.
", "rawMarkdown": "Currently using No.1 in my model.\n\nHaven't tried No.2, though there's a nicely upvoted kernel that does essentially this.\n\nNo.3 failed for me. I basically tried to follow Michael Jahrer's approach (from Porto) with a DAE but instead of using sparse features, I used an embedding on the categorical variable. Other changes were I also included some statistical features in the input which were not recreated in the reconstruction. Lastly, I implemented a custom Dense layer, which transposed the initial embedding matrix, essentially tying the weights and halving the model complexity. Maybe I bit off more than I can chew, or maybe I don't know what I'm doing at all. Trained with sparse cross entropy and the model learned... something because the final loss was about 20-25% it's initial loss. I only kept the categorical embeddings, threw them into my model as untrainable and ran it again. Loss dropped... but didn't beat--nor even come close--to my baseline.\n\nIf anyone else gives no.3 a whirl, please let me know how it goes. I tried training on train/train_a/test/test_a combined, for the available columns.\n\nAnd per @aquatic above, I will eventually try PL'ing.", "votes": 1 } ] }, { "id": 327933, "postDate": "2018-05-13T00:56:27.703Z", "content": "They could potentially be a nice source for pseudo-labeling / training set augmentation.
", "rawMarkdown": "They could potentially be a nice source for pseudo-labeling / training set augmentation.", "votes": 3 }, { "id": 329282, "postDate": "2018-05-16T06:27:56.497Z", "content": "I merged train_active data into train to try and create \"Duration of Ad\", \" Number of times an ad was activated\" type of features but realized none of the 'item_id' in 'train_period' appear in train. So how can we even use this data for training purposes ? ( Considering train_supplement and train_active do not have deal probability values)
\n\nCan someone tell me what I am missing?
\n\nBut we can definitely use train_active - 'title', 'item_description' data to generate better word embeddings.
", "rawMarkdown": "I merged train_active data into train to try and create \"Duration of Ad\", \" Number of times an ad was activated\" type of features but realized none of the 'item_id' in 'train_period' appear in train. So how can we even use this data for training purposes ? ( Considering train_supplement and train_active do not have deal probability values)\n\nCan someone tell me what I am missing?\n\nBut we can definitely use train_active - 'title', 'item_description' data to generate better word embeddings. ", "votes": 1 }, { "id": 328224, "postDate": "2018-05-13T18:08:58.837Z", "content": "I added a kernel for showing how to train a Word2Vec model from train_active.csv and another one for showing how to use a self-trained model and to compare with pre-trained Fasttext embeddings
", "rawMarkdown": "I added a kernel for showing [how to train a Word2Vec model][1] from train_active.csv and another one for showing how to [use a self-trained model][2] and to compare with pre-trained Fasttext embeddings\n\n\n [1]: https://www.kaggle.com/christofhenkel/using-train-active-for-training-word-embeddings\n [2]: https://www.kaggle.com/christofhenkel/self-trained-embeddings-starter-only-description", "votes": 2 }, { "id": 327797, "postDate": "2018-05-12T14:11:40.343Z", "content": "You can use your feature engineering on the combination of both sets.\nI was planning also on using test_active with the train_periods.csv to get more info about the time the ad remain available
", "rawMarkdown": "You can use your feature engineering on the combination of both sets.\nI was planning also on using test_active with the train_periods.csv to get more info about the time the ad remain available", "votes": 2 }, { "id": 344301, "postDate": "2018-06-17T14:56:26.170Z", "content": "You might want to check this kernel out:\nhttps://www.kaggle.com/bminixhofer/aggregated-features-lightgbm
", "rawMarkdown": "You might want to check this kernel out:\nhttps://www.kaggle.com/bminixhofer/aggregated-features-lightgbm" }, { "id": 343816, "postDate": "2018-06-16T07:45:20.073Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 343857, "postDate": "2018-06-16T11:10:52.537Z", "content": "int main() {\n return main();\n}
", "rawMarkdown": "int main() {\n return main();\n}", "votes": 1 }, { "id": 343900, "postDate": "2018-06-16T13:12:21.763Z", "content": "winrar.rar
", "rawMarkdown": "winrar.rar" }, { "id": 343913, "postDate": "2018-06-16T14:08:01.363Z", "content": "Sorry, changed it to the right link.
", "rawMarkdown": "Sorry, changed it to the right link." } ] } ], "comments": [ { "id": 327911, "author_name": "Dmytro Danevskyi", "author_url": "", "post_date": "2018-05-12T21:51:52.403000", "content": "Currently using No.1 in my model.
\n\nHaven't tried No.2, though there's a nicely upvoted kernel that does essentially this.
\n\nNo.3 failed for me. I basically tried to follow Michael Jahrer's approach (from Porto) with a DAE but instead of using sparse features, I used an embedding on the categorical variable. Other changes were I also included some statistical features in the input which were not recreated in the reconstruction. Lastly, I implemented a custom Dense layer, which transposed the initial embedding matrix, essentially tying the weights and halving the model complexity. Maybe I bit off more than I can chew, or maybe I don't know what I'm doing at all. Trained with sparse cross entropy and the model learned... something because the final loss was about 20-25% it's initial loss. I only kept the categorical embeddings, threw them into my model as untrainable and ran it again. Loss dropped... but didn't beat--nor even come close--to my baseline.
\n\nIf anyone else gives no.3 a whirl, please let me know how it goes. I tried training on train/train_a/test/test_a combined, for the available columns.
\n\nAnd per @aquatic above, I will eventually try PL'ing.
", "votes": 1, "replies": [] } ] }, { "id": 327933, "author_name": "Joe Eddy", "author_url": "", "post_date": "2018-05-13T00:56:27.703000", "content": "They could potentially be a nice source for pseudo-labeling / training set augmentation.
", "votes": 3, "replies": [] }, { "id": 329282, "author_name": "Shanth", "author_url": "", "post_date": "2018-05-16T06:27:56.497000", "content": "I merged train_active data into train to try and create \"Duration of Ad\", \" Number of times an ad was activated\" type of features but realized none of the 'item_id' in 'train_period' appear in train. So how can we even use this data for training purposes ? ( Considering train_supplement and train_active do not have deal probability values)
\n\nCan someone tell me what I am missing?
\n\nBut we can definitely use train_active - 'title', 'item_description' data to generate better word embeddings.
", "votes": 1, "replies": [] }, { "id": 328224, "author_name": "Dieter", "author_url": "", "post_date": "2018-05-13T18:08:58.837000", "content": "I added a kernel for showing how to train a Word2Vec model from train_active.csv and another one for showing how to use a self-trained model and to compare with pre-trained Fasttext embeddings
", "votes": 2, "replies": [] }, { "id": 327797, "author_name": "Antoine", "author_url": "", "post_date": "2018-05-12T14:11:40.343000", "content": "You can use your feature engineering on the combination of both sets.\nI was planning also on using test_active with the train_periods.csv to get more info about the time the ad remain available
", "votes": 2, "replies": [] }, { "id": 344301, "author_name": "Ian Dzindo", "author_url": "", "post_date": "2018-06-17T14:56:26.170000", "content": "You might want to check this kernel out:\nhttps://www.kaggle.com/bminixhofer/aggregated-features-lightgbm
", "votes": 0, "replies": [] }, { "id": 343816, "author_name": "", "author_url": "", "post_date": "2018-06-16T07:45:20.073000", "content": "", "votes": 0, "replies": [ { "id": 343857, "author_name": "عثمان", "author_url": "", "post_date": "2018-06-16T11:10:52.537000", "content": "int main() {\n return main();\n}
", "votes": 1, "replies": [] }, { "id": 343900, "author_name": "spongebob", "author_url": "", "post_date": "2018-06-16T13:12:21.763000", "content": "winrar.rar
", "votes": 0, "replies": [] }, { "id": 343913, "author_name": "Ian Dzindo", "author_url": "", "post_date": "2018-06-16T14:08:01.363000", "content": "Sorry, changed it to the right link.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "327605": "Since there is no lable train/test_active.csv. Text information can be collected as a corpus. What can be done to the others? Any advice will thank a lot.\n\n", "327911": " 1. They can be used to compute more accurate statistics on price/deal_probability, etc.\n 2. They can be used to create a better language model. Say, if you are using fasttext-like model with trainable embeddings, it might be beneficial to (unsupervisely) train these embeddings on texts from train/test active\n 3. You can make an autoencoder to capture dependencies between features. Since train/test active contain a LOT more data then train/test, it might lead to a good feature set\n\n", "327933": "They could potentially be a nice source for pseudo-labeling / training set augmentation.", "329282": "I merged train_active data into train to try and create \"Duration of Ad\", \" Number of times an ad was activated\" type of features but realized none of the 'item_id' in 'train_period' appear in train. So how can we even use this data for training purposes ? ( Considering train_supplement and train_active do not have deal probability values)\n\nCan someone tell me what I am missing?\n\nBut we can definitely use train_active - 'title', 'item_description' data to generate better word embeddings. ", "328224": "I added a kernel for showing [how to train a Word2Vec model][1] from train_active.csv and another one for showing how to [use a self-trained model][2] and to compare with pre-trained Fasttext embeddings\n\n\n [1]: https://www.kaggle.com/christofhenkel/using-train-active-for-training-word-embeddings\n [2]: https://www.kaggle.com/christofhenkel/self-trained-embeddings-starter-only-description", "327797": "You can use your feature engineering on the combination of both sets.\nI was planning also on using test_active with the train_periods.csv to get more info about the time the ad remain available", "344301": "You might want to check this kernel out:\nhttps://www.kaggle.com/bminixhofer/aggregated-features-lightgbm", "343816": "" } }