After taking some time to analyze the data, I can only conclude that this project either:
\n\nA) Is not put together properly.\nB) Is being purposely cryptic.
\n\nCan someone please confirm which one it is?
\n\nRight now, this data makes no sense. You have features included in the train file that are worthless and correlate to absolutely nothing. Furthermore, the expressed intent is to develop a new/better algorithm to classify. Well, having existing features negates this claim entirely, because unique features make or break any machine learning algorithm.
\n\nAs such, either the features are a ruse meant to add a cryptic layer to the challenge, or someone doesn't know what they're doing.
\n\nThe next issue is that it is stated that you can use whatever data you want basically, but you have very specific, predetermined targets for resulting scores. There's absolutely no explanation I can find of how these scores were determined. Again, either cryptic intentionally, or incompetence.
\n\nSo I ask, what's the point? Due to how this is structured, be it intentionally or not, all you're going to get out of this is \"Look at me, I can stuff arbitrary columns into some existing machine learning library I scarcely know anything about and get results. Yay!\" And that's exactly what we can see on the leaderboard already. \"LSTM->Stuff data-> result.\"
\n\nSurely existing engineers could sit here randomly stuffing data into random classification methods trying to match up with arbitrary predetermined outputs all the day long. So why throw away 50 grand for someone else to do the same thing?
\n\nI think these four issues need to be addressed at length to produce anything more than a showdown of what different existing open source libraries can do with this data.
", "messages": [ { "id": "430156", "postDate": "11/29/2018 23:17:36", "content": "After taking some time to analyze the data, I can only conclude that this project either:
\n\nA) Is not put together properly.\nB) Is being purposely cryptic.
\n\nCan someone please confirm which one it is?
\n\nRight now, this data makes no sense. You have features included in the train file that are worthless and correlate to absolutely nothing. Furthermore, the expressed intent is to develop a new/better algorithm to classify. Well, having existing features negates this claim entirely, because unique features make or break any machine learning algorithm.
\n\nAs such, either the features are a ruse meant to add a cryptic layer to the challenge, or someone doesn't know what they're doing.
\n\nThe next issue is that it is stated that you can use whatever data you want basically, but you have very specific, predetermined targets for resulting scores. There's absolutely no explanation I can find of how these scores were determined. Again, either cryptic intentionally, or incompetence.
\n\nSo I ask, what's the point? Due to how this is structured, be it intentionally or not, all you're going to get out of this is \"Look at me, I can stuff arbitrary columns into some existing machine learning library I scarcely know anything about and get results. Yay!\" And that's exactly what we can see on the leaderboard already. \"LSTM->Stuff data-> result.\"
\n\nSurely existing engineers could sit here randomly stuffing data into random classification methods trying to match up with arbitrary predetermined outputs all the day long. So why throw away 50 grand for someone else to do the same thing?
\n\nI think these four issues need to be addressed at length to produce anything more than a showdown of what different existing open source libraries can do with this data.
", "rawMarkdown": "After taking some time to analyze the data, I can only conclude that this project either:\n\nA) Is not put together properly.\nB) Is being purposely cryptic.\n\nCan someone please confirm which one it is? \n\nRight now, this data makes no sense. You have features included in the train file that are worthless and correlate to absolutely nothing. Furthermore, the expressed intent is to develop a new/better algorithm to classify. Well, having existing features negates this claim entirely, because unique features make or break any machine learning algorithm.\n\nAs such, either the features are a ruse meant to add a cryptic layer to the challenge, or someone doesn't know what they're doing.\n\nThe next issue is that it is stated that you can use whatever data you want basically, but you have very specific, predetermined targets for resulting scores. There's absolutely no explanation I can find of how these scores were determined. Again, either cryptic intentionally, or incompetence.\n\nSo I ask, what's the point? Due to how this is structured, be it intentionally or not, all you're going to get out of this is \"Look at me, I can stuff arbitrary columns into some existing machine learning library I scarcely know anything about and get results. Yay!\" And that's exactly what we can see on the leaderboard already. \"LSTM->Stuff data-> result.\"\n\nSurely existing engineers could sit here randomly stuffing data into random classification methods trying to match up with arbitrary predetermined outputs all the day long. So why throw away 50 grand for someone else to do the same thing?\n\nI think these four issues need to be addressed at length to produce anything more than a showdown of what different existing open source libraries can do with this data.", "votes": null }, { "id": "430926", "postDate": "12/01/2018 08:26:34", "content": "I think the main purpose of this competition is actually to find out if there is any underlying relationship between transactions behaviour and a customer's loyalty score. After knowing this relationship, ELO can predict a customer's loyalty towards ELO based on their historical transaction.
\n\nObviously you won't be able to find out how this 'Loyalty Score' is calculate because it is an internal business metric which takes into account both future spending and retention as main components. (Taken from https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/72993#430388) It is unlikely that they will reveal how this was calculated.
\n\nI do think that the overview description for the competition is not very clear about this purpose. And having the evaluation of RMSE doesn't mean that kagglers know what's the relationship between customer's loyalty and their respective transaction behaviours.
", "rawMarkdown": "I think the main purpose of this competition is actually to find out if there is any underlying relationship between transactions behaviour and a customer's loyalty score. After knowing this relationship, ELO can predict a customer's loyalty towards ELO based on their historical transaction.\n\nObviously you won't be able to find out how this 'Loyalty Score' is calculate because it is an internal business metric which takes into account both future spending and retention as main components. (Taken from https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/72993#430388) It is unlikely that they will reveal how this was calculated. \n\nI do think that the overview description for the competition is not very clear about this purpose. And having the evaluation of RMSE doesn't mean that kagglers know what's the relationship between customer's loyalty and their respective transaction behaviours.", "votes": null } ], "comments": [ { "id": 430926, "author_name": "teemingyi", "author_url": "", "post_date": "12/01/2018 08:26:34", "content": "I think the main purpose of this competition is actually to find out if there is any underlying relationship between transactions behaviour and a customer's loyalty score. After knowing this relationship, ELO can predict a customer's loyalty towards ELO based on their historical transaction.
\n\nObviously you won't be able to find out how this 'Loyalty Score' is calculate because it is an internal business metric which takes into account both future spending and retention as main components. (Taken from https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/72993#430388) It is unlikely that they will reveal how this was calculated.
\n\nI do think that the overview description for the competition is not very clear about this purpose. And having the evaluation of RMSE doesn't mean that kagglers know what's the relationship between customer's loyalty and their respective transaction behaviours.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "430156": "After taking some time to analyze the data, I can only conclude that this project either:\n\nA) Is not put together properly.\nB) Is being purposely cryptic.\n\nCan someone please confirm which one it is? \n\nRight now, this data makes no sense. You have features included in the train file that are worthless and correlate to absolutely nothing. Furthermore, the expressed intent is to develop a new/better algorithm to classify. Well, having existing features negates this claim entirely, because unique features make or break any machine learning algorithm.\n\nAs such, either the features are a ruse meant to add a cryptic layer to the challenge, or someone doesn't know what they're doing.\n\nThe next issue is that it is stated that you can use whatever data you want basically, but you have very specific, predetermined targets for resulting scores. There's absolutely no explanation I can find of how these scores were determined. Again, either cryptic intentionally, or incompetence.\n\nSo I ask, what's the point? Due to how this is structured, be it intentionally or not, all you're going to get out of this is \"Look at me, I can stuff arbitrary columns into some existing machine learning library I scarcely know anything about and get results. Yay!\" And that's exactly what we can see on the leaderboard already. \"LSTM->Stuff data-> result.\"\n\nSurely existing engineers could sit here randomly stuffing data into random classification methods trying to match up with arbitrary predetermined outputs all the day long. So why throw away 50 grand for someone else to do the same thing?\n\nI think these four issues need to be addressed at length to produce anything more than a showdown of what different existing open source libraries can do with this data.", "430926": "I think the main purpose of this competition is actually to find out if there is any underlying relationship between transactions behaviour and a customer's loyalty score. After knowing this relationship, ELO can predict a customer's loyalty towards ELO based on their historical transaction.\n\nObviously you won't be able to find out how this 'Loyalty Score' is calculate because it is an internal business metric which takes into account both future spending and retention as main components. (Taken from https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/72993#430388) It is unlikely that they will reveal how this was calculated. \n\nI do think that the overview description for the competition is not very clear about this purpose. And having the evaluation of RMSE doesn't mean that kagglers know what's the relationship between customer's loyalty and their respective transaction behaviours." }, "source": "meta" }