i just use the leak data as my additional training data , but my predict becomes worse, and the predict data is close to each other. I don't know why this happen.
", "messages": [ { "id": "681464", "postDate": "11/26/2019 06:31:30", "content": "i just use the leak data as my additional training data , but my predict becomes worse, and the predict data is close to each other. I don't know why this happen.
", "rawMarkdown": "i just use the leak data as my additional training data , but my predict becomes worse, and the predict data is close to each other. I don't know why this happen.", "votes": null }, { "id": "681488", "postDate": "11/26/2019 06:52:14", "content": "Which model are you using?\nIs the leak data added correctly during the pipeline or they have missing values on some features?
", "rawMarkdown": "Which model are you using?\nIs the leak data added correctly during the pipeline or they have missing values on some features?", "votes": null }, { "id": "681509", "postDate": "11/26/2019 07:19:16", "content": "just lightgbm, Five features has missing values, but just 17000.
", "rawMarkdown": "just lightgbm, Five features has missing values, but just 17000.", "votes": null }, { "id": "681516", "postDate": "11/26/2019 07:28:14", "content": "try to clean data (maybe you have other error in the pipeline).\nI used leak data and i observed better model and better cv estimation.
", "rawMarkdown": "try to clean data (maybe you have other error in the pipeline).\nI used leak data and i observed better model and better cv estimation.", "votes": null }, { "id": "681519", "postDate": "11/26/2019 07:34:38", "content": "Although I do not know a lot , this is one possibility. Meter readings of the same building with different years are tend to be high correlation, but the other buildings might tend to get poor correlation with that building. If leaked data is overly fitted to the building itself. The model could lose generalization of the overall trends. I have some doubts augmenting leaked data is helpful to predict the others also. Meter readings of each building are having quiet different behavior? Then adding meter readings of another year is helpful?
", "rawMarkdown": "Although I do not know a lot , this is one possibility. Meter readings of the same building with different years are tend to be high correlation, but the other buildings might tend to get poor correlation with that building. If leaked data is overly fitted to the building itself. The model could lose generalization of the overall trends. I have some doubts augmenting leaked data is helpful to predict the others also. Meter readings of each building are having quiet different behavior? Then adding meter readings of another year is helpful?", "votes": null }, { "id": "681530", "postDate": "11/26/2019 07:50:03", "content": "What I thought about the leaked data at the first time. It could be used for validation set set not train set. This might be highly useful , we could find some nice spot penetrating all meter reading trends in years only if it could be found. I do not use it though anyways.
", "rawMarkdown": "What I thought about the leaked data at the first time. It could be used for validation set set not train set. This might be highly useful , we could find some nice spot penetrating all meter reading trends in years only if it could be found. I do not use it though anyways.", "votes": null }, { "id": "681535", "postDate": "11/26/2019 07:55:53", "content": "I think your reply maybe right .My predict data has many same number among different building. Maybe adding leak data makes model overfitted. And i think whether adding leak data make the distributions of training data far different from test data.However i also find in public kernel, someone using leak data has a good result.And i don't know how they do that.
", "rawMarkdown": "I think your reply maybe right .My predict data has many same number among different building. Maybe adding leak data makes model overfitted. And i think whether adding leak data make the distributions of training data far different from test data.However i also find in public kernel, someone using leak data has a good result.And i don't know how they do that.", "votes": null }, { "id": "681537", "postDate": "11/26/2019 07:58:00", "content": "Yes that could make your train set leaked. ;) I did not quiet see that kernel, is that kernel predicting itself with leaked data of that building?
", "rawMarkdown": "Yes that could make your train set leaked. ;) I did not quiet see that kernel, is that kernel predicting itself with leaked data of that building?", "votes": null }, { "id": "681547", "postDate": "11/26/2019 08:15:22", "content": "My memory is wrong. ;) i saw some says that he uses leak data mixing with training data in discussion rather than public kernel.
", "rawMarkdown": "My memory is wrong. ;) i saw some says that he uses leak data mixing with training data in discussion rather than public kernel.", "votes": null } ], "comments": [ { "id": 681488, "author_name": "stenford23", "author_url": "", "post_date": "11/26/2019 06:52:14", "content": "Which model are you using?\nIs the leak data added correctly during the pipeline or they have missing values on some features?
", "votes": null, "replies": [] }, { "id": 681509, "author_name": "jacobfrey", "author_url": "", "post_date": "11/26/2019 07:19:16", "content": "just lightgbm, Five features has missing values, but just 17000.
", "votes": null, "replies": [ { "id": 681516, "author_name": "stenford23", "author_url": "", "post_date": "11/26/2019 07:28:14", "content": "try to clean data (maybe you have other error in the pipeline).\nI used leak data and i observed better model and better cv estimation.
", "votes": null, "replies": [] } ] }, { "id": 681519, "author_name": "hanjoonchoe", "author_url": "", "post_date": "11/26/2019 07:34:38", "content": "Although I do not know a lot , this is one possibility. Meter readings of the same building with different years are tend to be high correlation, but the other buildings might tend to get poor correlation with that building. If leaked data is overly fitted to the building itself. The model could lose generalization of the overall trends. I have some doubts augmenting leaked data is helpful to predict the others also. Meter readings of each building are having quiet different behavior? Then adding meter readings of another year is helpful?
", "votes": null, "replies": [ { "id": 681530, "author_name": "hanjoonchoe", "author_url": "", "post_date": "11/26/2019 07:50:03", "content": "What I thought about the leaked data at the first time. It could be used for validation set set not train set. This might be highly useful , we could find some nice spot penetrating all meter reading trends in years only if it could be found. I do not use it though anyways.
", "votes": null, "replies": [] } ] }, { "id": 681535, "author_name": "jacobfrey", "author_url": "", "post_date": "11/26/2019 07:55:53", "content": "I think your reply maybe right .My predict data has many same number among different building. Maybe adding leak data makes model overfitted. And i think whether adding leak data make the distributions of training data far different from test data.However i also find in public kernel, someone using leak data has a good result.And i don't know how they do that.
", "votes": null, "replies": [ { "id": 681537, "author_name": "hanjoonchoe", "author_url": "", "post_date": "11/26/2019 07:58:00", "content": "Yes that could make your train set leaked. ;) I did not quiet see that kernel, is that kernel predicting itself with leaked data of that building?
", "votes": null, "replies": [] } ] }, { "id": 681547, "author_name": "jacobfrey", "author_url": "", "post_date": "11/26/2019 08:15:22", "content": "My memory is wrong. ;) i saw some says that he uses leak data mixing with training data in discussion rather than public kernel.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "681464": "i just use the leak data as my additional training data , but my predict becomes worse, and the predict data is close to each other. I don't know why this happen.", "681488": "Which model are you using?\nIs the leak data added correctly during the pipeline or they have missing values on some features?", "681509": "just lightgbm, Five features has missing values, but just 17000.", "681516": "try to clean data (maybe you have other error in the pipeline).\nI used leak data and i observed better model and better cv estimation.", "681519": "Although I do not know a lot , this is one possibility. Meter readings of the same building with different years are tend to be high correlation, but the other buildings might tend to get poor correlation with that building. If leaked data is overly fitted to the building itself. The model could lose generalization of the overall trends. I have some doubts augmenting leaked data is helpful to predict the others also. Meter readings of each building are having quiet different behavior? Then adding meter readings of another year is helpful?", "681530": "What I thought about the leaked data at the first time. It could be used for validation set set not train set. This might be highly useful , we could find some nice spot penetrating all meter reading trends in years only if it could be found. I do not use it though anyways.", "681535": "I think your reply maybe right .My predict data has many same number among different building. Maybe adding leak data makes model overfitted. And i think whether adding leak data make the distributions of training data far different from test data.However i also find in public kernel, someone using leak data has a good result.And i don't know how they do that.", "681537": "Yes that could make your train set leaked. ;) I did not quiet see that kernel, is that kernel predicting itself with leaked data of that building?", "681547": "My memory is wrong. ;) i saw some says that he uses leak data mixing with training data in discussion rather than public kernel." }, "source": "meta" }