Hello Everyone,
\n\nHere is my approach on this problem:
\n\nWould greatly appreciate any responses to this initial thoughts. Since dealing with this data is computationally expensive, i thought its better to take opinions from Kaggle Masters here :) Please excuse me if my approach sounds too naive.
\n\nThanks!
", "messages": [ { "id": "126253", "postDate": "07/07/2016 05:51:09", "content": "Hello Everyone,
\n\nHere is my approach on this problem:
\n\nWould greatly appreciate any responses to this initial thoughts. Since dealing with this data is computationally expensive, i thought its better to take opinions from Kaggle Masters here :) Please excuse me if my approach sounds too naive.
\n\nThanks!
", "rawMarkdown": "Hello Everyone,\r\n\r\nHere is my approach on this problem:\r\n\r\n1. Since there are too many classes for few dimensions like Agencia_ID, Producto_ID; I was thinking to create tiers/rank for the classes based on the overall Demand. This will replace unique IDs with tiers (maybe 20 or 30)\r\n2. Split the training data into train and split in 70:30 ratio\r\n3. Adopt random forest to start with\r\n\r\nWould greatly appreciate any responses to this initial thoughts. Since dealing with this data is computationally expensive, i thought its better to take opinions from Kaggle Masters here :) Please excuse me if my approach sounds too naive. \r\n\r\nThanks!", "votes": null }, { "id": "126282", "postDate": "07/07/2016 10:32:01", "content": "There is only one way to know for sure: just do it. Look at the resulting LB score, and at your local validation score.
\n\nNext thing IMHO is to check that each time your local score improves then so does the LB score.
\n\nOnce you have this, then you can start real work.
", "rawMarkdown": "There is only one way to know for sure: just do it. Look at the resulting LB score, and at your local validation score.\r\n\r\nNext thing IMHO is to check that each time your local score improves then so does the LB score.\r\n\r\nOnce you have this, then you can start real work.", "votes": null }, { "id": "126295", "postDate": "07/07/2016 14:26:03", "content": "Here is something that I found out: the IDs carry some information with them, because they are not random. So the distance between them may be a valid info about how old a client or product is with respect to others. If you just drop them and work with tiers you may be loosing that info. Also, by ploting a simple histogram of the IDs you can see some interesting patterns.
", "rawMarkdown": "Here is something that I found out: the IDs carry some information with them, because they are not random. So the distance between them may be a valid info about how old a client or product is with respect to others. If you just drop them and work with tiers you may be loosing that info. Also, by ploting a simple histogram of the IDs you can see some interesting patterns.", "votes": null }, { "id": "126389", "postDate": "07/08/2016 04:57:21", "content": "Thank you CPMP and Regis for your valuable comments. I will try my current methodology and will reach out if I need further help.
", "rawMarkdown": "Thank you CPMP and Regis for your valuable comments. I will try my current methodology and will reach out if I need further help.", "votes": null } ], "comments": [ { "id": 126282, "author_name": "cpmpml", "author_url": "", "post_date": "07/07/2016 10:32:01", "content": "There is only one way to know for sure: just do it. Look at the resulting LB score, and at your local validation score.
\n\nNext thing IMHO is to check that each time your local score improves then so does the LB score.
\n\nOnce you have this, then you can start real work.
", "votes": null, "replies": [] }, { "id": 126295, "author_name": "regisely", "author_url": "", "post_date": "07/07/2016 14:26:03", "content": "Here is something that I found out: the IDs carry some information with them, because they are not random. So the distance between them may be a valid info about how old a client or product is with respect to others. If you just drop them and work with tiers you may be loosing that info. Also, by ploting a simple histogram of the IDs you can see some interesting patterns.
", "votes": null, "replies": [] }, { "id": 126389, "author_name": "", "author_url": "", "post_date": "07/08/2016 04:57:21", "content": "Thank you CPMP and Regis for your valuable comments. I will try my current methodology and will reach out if I need further help.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "126253": "Hello Everyone,\r\n\r\nHere is my approach on this problem:\r\n\r\n1. Since there are too many classes for few dimensions like Agencia_ID, Producto_ID; I was thinking to create tiers/rank for the classes based on the overall Demand. This will replace unique IDs with tiers (maybe 20 or 30)\r\n2. Split the training data into train and split in 70:30 ratio\r\n3. Adopt random forest to start with\r\n\r\nWould greatly appreciate any responses to this initial thoughts. Since dealing with this data is computationally expensive, i thought its better to take opinions from Kaggle Masters here :) Please excuse me if my approach sounds too naive. \r\n\r\nThanks!", "126282": "There is only one way to know for sure: just do it. Look at the resulting LB score, and at your local validation score.\r\n\r\nNext thing IMHO is to check that each time your local score improves then so does the LB score.\r\n\r\nOnce you have this, then you can start real work.", "126295": "Here is something that I found out: the IDs carry some information with them, because they are not random. So the distance between them may be a valid info about how old a client or product is with respect to others. If you just drop them and work with tiers you may be loosing that info. Also, by ploting a simple histogram of the IDs you can see some interesting patterns.", "126389": "Thank you CPMP and Regis for your valuable comments. I will try my current methodology and will reach out if I need further help." }, "source": "meta" }