Generating a large number of candidates, such as a co-visitation matrix, is often expected to result in an out-of-memory error.
\nTherefore, I am considering negative sampling of some unnecessary candidates or sessions.
\nAs a kaggle beginner, I can only think of the followings…
\nDo you have any good ideas?
", "messages": [ { "id": "2071703", "postDate": "12/21/2022 09:33:29", "content": "Generating a large number of candidates, such as a co-visitation matrix, is often expected to result in an out-of-memory error.
\nTherefore, I am considering negative sampling of some unnecessary candidates or sessions.
\nAs a kaggle beginner, I can only think of the followings…
\nDo you have any good ideas?
", "rawMarkdown": "Generating a large number of candidates, such as a co-visitation matrix, is often expected to result in an out-of-memory error.\n\nTherefore, I am considering negative sampling of some unnecessary candidates or sessions.\n\nAs a kaggle beginner, I can only think of the followings...\n\n- Remove sessions with too few logs\n- Randomly reduce the number of sessions.\n- Reduce the number of batches per study\n\nDo you have any good ideas?", "votes": null }, { "id": "2071752", "postDate": "12/21/2022 10:49:13", "content": "Otto's recent approach discusses about negative sampling (although the original video is in German).
\nAs far as I know the most popular and recent approach of negative sampling is that used for GRU4Rec & GRU4Rec+. Where they sample negative samples out of batch. Sampled Softmax Loss is also an approach that reduces memory consumption.
However, I think most of current approach requires huge amount of GPU memory to achieve a decent score on a huge number of item & users.
\nIf you have not enough memory, I think candidate rerank model shared in this discussion is most promising ones.
", "rawMarkdown": "[Otto's recent approach] discusses about negative sampling (although the original video is in German).\nAs far as I know the most popular and recent approach of negative sampling is that used for GRU4Rec & GRU4Rec+. Where they sample negative samples out of batch. Sampled Softmax Loss is also an approach that reduces memory consumption.\n\nHowever, I think most of current approach requires huge amount of GPU memory to achieve a decent score on a huge number of item & users.\n\nIf you have not enough memory, I think [candidate rerank model shared in this discussion] is most promising ones.\n\n[Otto's recent approach]: https://www.kaggle.com/competitions/otto-recommender-system/discussion/373224\n[candidate rerank model shared in this discussion]: https://www.kaggle.com/competitions/otto-recommender-system/discussion/364721", "votes": null }, { "id": "2073685", "postDate": "12/23/2022 09:19:29", "content": "Build your validation scheme, and try all of it.
\nFYI, I'm using random.
Does it simply work? Did you try some heavy searches (hyperparameters for example) based on it? It didn't break?
", "rawMarkdown": "Does it simply work? Did you try some heavy searches (hyperparameters for example) based on it? It didn't break?", "votes": null } ], "comments": [ { "id": 2071752, "author_name": "tatamikenn", "author_url": "", "post_date": "12/21/2022 10:49:13", "content": "Otto's recent approach discusses about negative sampling (although the original video is in German).
\nAs far as I know the most popular and recent approach of negative sampling is that used for GRU4Rec & GRU4Rec+. Where they sample negative samples out of batch. Sampled Softmax Loss is also an approach that reduces memory consumption.
However, I think most of current approach requires huge amount of GPU memory to achieve a decent score on a huge number of item & users.
\nIf you have not enough memory, I think candidate rerank model shared in this discussion is most promising ones.
", "votes": null, "replies": [] }, { "id": 2073685, "author_name": "onodera", "author_url": "", "post_date": "12/23/2022 09:19:29", "content": "Build your validation scheme, and try all of it.
\nFYI, I'm using random.
Does it simply work? Did you try some heavy searches (hyperparameters for example) based on it? It didn't break?
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2071703": "Generating a large number of candidates, such as a co-visitation matrix, is often expected to result in an out-of-memory error.\n\nTherefore, I am considering negative sampling of some unnecessary candidates or sessions.\n\nAs a kaggle beginner, I can only think of the followings...\n\n- Remove sessions with too few logs\n- Randomly reduce the number of sessions.\n- Reduce the number of batches per study\n\nDo you have any good ideas?", "2071752": "[Otto's recent approach] discusses about negative sampling (although the original video is in German).\nAs far as I know the most popular and recent approach of negative sampling is that used for GRU4Rec & GRU4Rec+. Where they sample negative samples out of batch. Sampled Softmax Loss is also an approach that reduces memory consumption.\n\nHowever, I think most of current approach requires huge amount of GPU memory to achieve a decent score on a huge number of item & users.\n\nIf you have not enough memory, I think [candidate rerank model shared in this discussion] is most promising ones.\n\n[Otto's recent approach]: https://www.kaggle.com/competitions/otto-recommender-system/discussion/373224\n[candidate rerank model shared in this discussion]: https://www.kaggle.com/competitions/otto-recommender-system/discussion/364721", "2073685": "Build your validation scheme, and try all of it.\nFYI, I'm using random.", "2076434": "Does it simply work? Did you try some heavy searches (hyperparameters for example) based on it? It didn't break?" }, "source": "meta" }