I have checked some kernels. It seems many of them just mix all the stores and items together and then feed these data (after some feature engineering) into a model like XGBoost. But I am confused and hope to get some help here.
\n\nDoes it make sense by mixing them up? In a decision tree based model, the *store_nbr* and *item_nbr* may be used to split nodes. If so, why don't we just train a model for each (store, item) pair? After all, we are required to predict for each (store, item) separately.
", "messages": [ { "id": "266760", "postDate": "01/09/2018 16:49:48", "content": "I have checked some kernels. It seems many of them just mix all the stores and items together and then feed these data (after some feature engineering) into a model like XGBoost. But I am confused and hope to get some help here.
\n\nDoes it make sense by mixing them up? In a decision tree based model, the *store_nbr* and *item_nbr* may be used to split nodes. If so, why don't we just train a model for each (store, item) pair? After all, we are required to predict for each (store, item) separately.
", "rawMarkdown": "I have checked some kernels. It seems many of them just mix all the stores and items together and then feed these data (after some feature engineering) into a model like XGBoost. But I am confused and hope to get some help here.\n\nDoes it make sense by mixing them up? In a decision tree based model, the *store_nbr* and *item_nbr* may be used to split nodes. If so, why don't we just train a model for each **(store, item)** pair? After all, we are required to predict for each **(store, item)** separately.", "votes": null }, { "id": "266796", "postDate": "01/09/2018 19:33:42", "content": "In my opinion, if you use store_nbr and item_nbr to split nodes, it would be too specific and thus there would be no enough training data to train such a specific (store_nbr, item_nbr) nodes. In other word, mixing them up can make the model more generative.
", "rawMarkdown": "In my opinion, if you use store_nbr and item_nbr to split nodes, it would be too specific and thus there would be no enough training data to train such a specific (store_nbr, item_nbr) nodes. In other word, mixing them up can make the model more generative.", "votes": null }, { "id": "266881", "postDate": "01/10/2018 03:55:19", "content": "Hi, ZedYeung. Thanks for your reply. Yes, you are right. If we split the training data according to (store_nbr, item_nbr), there will be only around 350 samples on average for each (store_nbr, item_nbr). However, it is still a little confusing. In traditional time series analysis like the ARIMA model, we need to split these data, each time series for each (store_nbr, item_nbr), right?
", "rawMarkdown": "Hi, ZedYeung. Thanks for your reply. Yes, you are right. If we split the training data according to (store_nbr, item_nbr), there will be only around 350 samples on average for each (store_nbr, item_nbr). However, it is still a little confusing. In traditional time series analysis like the ARIMA model, we need to split these data, each time series for each (store_nbr, item_nbr), right?", "votes": null }, { "id": "267191", "postDate": "01/10/2018 20:06:13", "content": "I have this concerns as well. I was thinking about if we have a separate model for each (store, item) pair, for example one LSTM for each pair's time series, we would end up having a large number of LSTMs. Just feel it's not right to have that many models.
", "rawMarkdown": "I have this concerns as well. I was thinking about if we have a separate model for each (store, item) pair, for example one LSTM for each pair's time series, we would end up having a large number of LSTMs. Just feel it's not right to have that many models.", "votes": null }, { "id": "267479", "postDate": "01/11/2018 13:16:34", "content": "I tried this, but results improved only for small number of pairs with high average demand. I think we just haven't enough data for only one pair in general case.
", "rawMarkdown": "I tried this, but results improved only for small number of pairs with high average demand. I think we just haven't enough data for only one pair in general case.", "votes": null } ], "comments": [ { "id": 266796, "author_name": "yedthon", "author_url": "", "post_date": "01/09/2018 19:33:42", "content": "In my opinion, if you use store_nbr and item_nbr to split nodes, it would be too specific and thus there would be no enough training data to train such a specific (store_nbr, item_nbr) nodes. In other word, mixing them up can make the model more generative.
", "votes": null, "replies": [ { "id": 266881, "author_name": "shuhuagao", "author_url": "", "post_date": "01/10/2018 03:55:19", "content": "Hi, ZedYeung. Thanks for your reply. Yes, you are right. If we split the training data according to (store_nbr, item_nbr), there will be only around 350 samples on average for each (store_nbr, item_nbr). However, it is still a little confusing. In traditional time series analysis like the ARIMA model, we need to split these data, each time series for each (store_nbr, item_nbr), right?
", "votes": null, "replies": [] } ] }, { "id": 267191, "author_name": "bdeng3", "author_url": "", "post_date": "01/10/2018 20:06:13", "content": "I have this concerns as well. I was thinking about if we have a separate model for each (store, item) pair, for example one LSTM for each pair's time series, we would end up having a large number of LSTMs. Just feel it's not right to have that many models.
", "votes": null, "replies": [] }, { "id": 267479, "author_name": "ilyalos", "author_url": "", "post_date": "01/11/2018 13:16:34", "content": "I tried this, but results improved only for small number of pairs with high average demand. I think we just haven't enough data for only one pair in general case.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "266760": "I have checked some kernels. It seems many of them just mix all the stores and items together and then feed these data (after some feature engineering) into a model like XGBoost. But I am confused and hope to get some help here.\n\nDoes it make sense by mixing them up? In a decision tree based model, the *store_nbr* and *item_nbr* may be used to split nodes. If so, why don't we just train a model for each **(store, item)** pair? After all, we are required to predict for each **(store, item)** separately.", "266796": "In my opinion, if you use store_nbr and item_nbr to split nodes, it would be too specific and thus there would be no enough training data to train such a specific (store_nbr, item_nbr) nodes. In other word, mixing them up can make the model more generative.", "266881": "Hi, ZedYeung. Thanks for your reply. Yes, you are right. If we split the training data according to (store_nbr, item_nbr), there will be only around 350 samples on average for each (store_nbr, item_nbr). However, it is still a little confusing. In traditional time series analysis like the ARIMA model, we need to split these data, each time series for each (store_nbr, item_nbr), right?", "267191": "I have this concerns as well. I was thinking about if we have a separate model for each (store, item) pair, for example one LSTM for each pair's time series, we would end up having a large number of LSTMs. Just feel it's not right to have that many models.", "267479": "I tried this, but results improved only for small number of pairs with high average demand. I think we just haven't enough data for only one pair in general case." }, "source": "meta" }