Why, concatenating all transactions doesn't improve the model, but individually is better?
", "messages": [ { "id": "458546", "postDate": "01/19/2019 22:29:50", "content": "Why, concatenating all transactions doesn't improve the model, but individually is better?
", "rawMarkdown": "Why, concatenating all transactions doesn't improve the model, but individually is better?", "votes": null }, { "id": "458657", "postDate": "01/20/2019 07:42:21", "content": "Cuz they are different. Historical transactions are transaction which were made before any discounts(promotion) , while new_merchant transactions are new transactions after discount is given, and only new merchants are included.
", "rawMarkdown": "Cuz they are different. Historical transactions are transaction which were made before any discounts(promotion) , while new_merchant transactions are new transactions after discount is given, and only new merchants are included.", "votes": null }, { "id": "458667", "postDate": "01/20/2019 08:23:17", "content": "@Salamat, can you please explain out what exactly new_merchant transactions are in simple terms please ? I dind't quite understand the way it's mentioned in discussion forum and problem statement . \nAlso, does that mean, if any user has made transactions in the period ( covered by new_merchants) with older merchants it is not listed anywhere ( neither historical_transactions nor new_merchants) ?
", "rawMarkdown": "Salamat, can you please explain out what exactly new_merchant transactions are in simple terms please ? I dind't quite understand the way it's mentioned in discussion forum and problem statement . \nAlso, does that mean, if any user has made transactions in the period ( covered by new_merchants) with older merchants it is not listed anywhere ( neither historical_transactions nor new_merchants) ?", "votes": null }, { "id": "458855", "postDate": "01/20/2019 16:36:59", "content": "May I know how you get this conclusion? TQ
", "rawMarkdown": "May I know how you get this conclusion? TQ", "votes": null }, { "id": "458975", "postDate": "01/20/2019 23:43:24", "content": "i made a model approach according to the fact that new merchants transactions is an evaluation perdiod and still there are incosistencies on the training dataset. For example card_id=C_ID_29d6a47e35, has not transactions at all on the new merchants file and has a loyalty score of 2.584628, which is pretty good score. This card has it's latest transaction really late, on 11/6/2017
\n\n2 months now, i try to understand what is the target, but is seems that the competition goal is to find the \"what is the target\" :D
\n\nThe opposite happens for card=C_ID_3f7ba15ec2, where has target -17.49263 and on the new merchants file, this card visited 7 different merchants, and on the one had a purchase amount of 1.5, which was the specific merchant's highest transaction amount on the new merchants file(M_ID_f40eec7eb5).... i believe that it is not clear what we are predicting and how the data were simulated. Maybe we are trying to predict noise.... who knows.....
", "rawMarkdown": "i made a model approach according to the fact that new merchants transactions is an evaluation perdiod and still there are incosistencies on the training dataset. For example card_id=C_ID_29d6a47e35, has not transactions at all on the new merchants file and has a loyalty score of 2.584628, which is pretty good score. This card has it's latest transaction really late, on 11/6/2017\n\n2 months now, i try to understand what is the target, but is seems that the competition goal is to find the \"what is the target\" :D \n\nThe opposite happens for card=C_ID_3f7ba15ec2, where has target -17.49263 and on the new merchants file, this card visited 7 different merchants, and on the one had a purchase amount of 1.5, which was the specific merchant's highest transaction amount on the new merchants file(M_ID_f40eec7eb5).... i believe that it is not clear what we are predicting and how the data were simulated. Maybe we are trying to predict noise.... who knows.....", "votes": null }, { "id": "459435", "postDate": "01/21/2019 18:16:56", "content": "I think new merchant transactions should help in building features that capture degree of loyalty to the core merchant group. I call merchants in the historic transaction a legacy group and merchants in new transactions a diviant group. From that I have built features that rank high on feature importance scale, but it did not translate to the higher LB score... I hope for the time being :)
", "rawMarkdown": "I think new merchant transactions should help in building features that capture degree of loyalty to the core merchant group. I call merchants in the historic transaction a legacy group and merchants in new transactions a diviant group. From that I have built features that rank high on feature importance scale, but it did not translate to the higher LB score... I hope for the time being :)", "votes": null }, { "id": "459707", "postDate": "01/22/2019 08:05:57", "content": "skw1990:\nYou can check if you look at new merchant transaction's merchants they will be all unique and almost all of them not present history transactions. There will be some examples where merchants in historical transaction also appear in new merchant transactions but number is very low ~1 (same merchant) for each card_id.
", "rawMarkdown": "skw1990:\nYou can check if you look at new merchant transaction's merchants they will be all unique and almost all of them not present history transactions. There will be some examples where merchants in historical transaction also appear in new merchant transactions but number is very low ~1 (same merchant) for each card_id.", "votes": null }, { "id": "459710", "postDate": "01/22/2019 08:15:25", "content": "Dimosthenis Karaflos\nI am not really sure whether positive target is good or not (maybe in fact smaller is better). My understanding is that we are predicting customer loyality score based on new merchant. However, there are cases that you mentioned. I am not sure how to explain them.
", "rawMarkdown": "Dimosthenis Karaflos\nI am not really sure whether positive target is good or not (maybe in fact smaller is better). My understanding is that we are predicting customer loyality score based on new merchant. However, there are cases that you mentioned. I am not sure how to explain them.", "votes": null }, { "id": "459712", "postDate": "01/22/2019 08:20:19", "content": "Sorry, I can't make it simpler. Try to explore data look at the dates and unique merchats in new merchant transaction. Also, make sure that you read overview and explanations of data.
", "rawMarkdown": "Sorry, I can't make it simpler. Try to explore data look at the dates and unique merchats in new merchant transaction. Also, make sure that you read overview and explanations of data.", "votes": null }, { "id": "459713", "postDate": "01/22/2019 08:24:07", "content": "@salamat\nIndeed, the data seem irrelevant to the target and the host does not give clear explanations about the target and the normalization process.
", "rawMarkdown": "salamat\nIndeed, the data seem irrelevant to the target and the host does not give clear explanations about the target and the normalization process.", "votes": null }, { "id": "459718", "postDate": "01/22/2019 08:37:14", "content": "I think what they are trying to see whether we can predict customer loyalty score for new merchants based on behavior of customer .From historical transactions one might understand about customers preference and behavior. This might also true for new merchant but I think we need it to find some correlations(with historical transactions) which I haven't really tried yet. So, problem is not straightforward we don't really know which features are relevant to customer loyalty for new merchants. I believe problem would be much simpler if we predict customer loyalty score for \"old\" merchant where more information is available(historical transactions).
", "rawMarkdown": "I think what they are trying to see whether we can predict customer loyalty score for new merchants based on behavior of customer .From historical transactions one might understand about customers preference and behavior. This might also true for new merchant but I think we need it to find some correlations(with historical transactions) which I haven't really tried yet. So, problem is not straightforward we don't really know which features are relevant to customer loyalty for new merchants. I believe problem would be much simpler if we predict customer loyalty score for \"old\" merchant where more information is available(historical transactions).", "votes": null }, { "id": "459741", "postDate": "01/22/2019 09:18:51", "content": "In fact, i tried to read about ELO, and i think that the recommendations are not personalized. I believe that it is like discounts of 20% for example, on a specific merchant. But this discount is available for all the cardholders, so collaboration filtering, FMs,MFdo not seem appropriate methodology if this is the case.
", "rawMarkdown": "In fact, i tried to read about ELO, and i think that the recommendations are not personalized. I believe that it is like discounts of 20% for example, on a specific merchant. But this discount is available for all the cardholders, so collaboration filtering, FMs,MFdo not seem appropriate methodology if this is the case.", "votes": null }, { "id": "463249", "postDate": "01/29/2019 17:14:28", "content": "Thanks.
\n\nBased on my observations, at card level, merchants in old & new transactions are not overlapped; Also, some customers do not have transactions, yet appear to have high score....\nActually I still do not comprehend the target. The loyalty score here is target for merchant or target for customers.....puzzling haha
", "rawMarkdown": "Thanks.\n\nBased on my observations, at card level, merchants in old & new transactions are not overlapped; Also, some customers do not have transactions, yet appear to have high score....\nActually I still do not comprehend the target. The loyalty score here is target for merchant or target for customers.....puzzling haha", "votes": null } ], "comments": [ { "id": 458657, "author_name": "salamatburj", "author_url": "", "post_date": "01/20/2019 07:42:21", "content": "Cuz they are different. Historical transactions are transaction which were made before any discounts(promotion) , while new_merchant transactions are new transactions after discount is given, and only new merchants are included.
", "votes": null, "replies": [ { "id": 458855, "author_name": "skw1990", "author_url": "", "post_date": "01/20/2019 16:36:59", "content": "May I know how you get this conclusion? TQ
", "votes": null, "replies": [] }, { "id": 458975, "author_name": "dkaraflos", "author_url": "", "post_date": "01/20/2019 23:43:24", "content": "i made a model approach according to the fact that new merchants transactions is an evaluation perdiod and still there are incosistencies on the training dataset. For example card_id=C_ID_29d6a47e35, has not transactions at all on the new merchants file and has a loyalty score of 2.584628, which is pretty good score. This card has it's latest transaction really late, on 11/6/2017
\n\n2 months now, i try to understand what is the target, but is seems that the competition goal is to find the \"what is the target\" :D
\n\nThe opposite happens for card=C_ID_3f7ba15ec2, where has target -17.49263 and on the new merchants file, this card visited 7 different merchants, and on the one had a purchase amount of 1.5, which was the specific merchant's highest transaction amount on the new merchants file(M_ID_f40eec7eb5).... i believe that it is not clear what we are predicting and how the data were simulated. Maybe we are trying to predict noise.... who knows.....
", "votes": null, "replies": [] }, { "id": 459707, "author_name": "salamatburj", "author_url": "", "post_date": "01/22/2019 08:05:57", "content": "skw1990:\nYou can check if you look at new merchant transaction's merchants they will be all unique and almost all of them not present history transactions. There will be some examples where merchants in historical transaction also appear in new merchant transactions but number is very low ~1 (same merchant) for each card_id.
", "votes": null, "replies": [] }, { "id": 459710, "author_name": "salamatburj", "author_url": "", "post_date": "01/22/2019 08:15:25", "content": "Dimosthenis Karaflos\nI am not really sure whether positive target is good or not (maybe in fact smaller is better). My understanding is that we are predicting customer loyality score based on new merchant. However, there are cases that you mentioned. I am not sure how to explain them.
", "votes": null, "replies": [] }, { "id": 459713, "author_name": "dkaraflos", "author_url": "", "post_date": "01/22/2019 08:24:07", "content": "@salamat\nIndeed, the data seem irrelevant to the target and the host does not give clear explanations about the target and the normalization process.
", "votes": null, "replies": [] }, { "id": 459718, "author_name": "salamatburj", "author_url": "", "post_date": "01/22/2019 08:37:14", "content": "I think what they are trying to see whether we can predict customer loyalty score for new merchants based on behavior of customer .From historical transactions one might understand about customers preference and behavior. This might also true for new merchant but I think we need it to find some correlations(with historical transactions) which I haven't really tried yet. So, problem is not straightforward we don't really know which features are relevant to customer loyalty for new merchants. I believe problem would be much simpler if we predict customer loyalty score for \"old\" merchant where more information is available(historical transactions).
", "votes": null, "replies": [] }, { "id": 459741, "author_name": "dkaraflos", "author_url": "", "post_date": "01/22/2019 09:18:51", "content": "In fact, i tried to read about ELO, and i think that the recommendations are not personalized. I believe that it is like discounts of 20% for example, on a specific merchant. But this discount is available for all the cardholders, so collaboration filtering, FMs,MFdo not seem appropriate methodology if this is the case.
", "votes": null, "replies": [] }, { "id": 463249, "author_name": "skw1990", "author_url": "", "post_date": "01/29/2019 17:14:28", "content": "Thanks.
\n\nBased on my observations, at card level, merchants in old & new transactions are not overlapped; Also, some customers do not have transactions, yet appear to have high score....\nActually I still do not comprehend the target. The loyalty score here is target for merchant or target for customers.....puzzling haha
", "votes": null, "replies": [] } ] }, { "id": 458667, "author_name": "vinuthanayak", "author_url": "", "post_date": "01/20/2019 08:23:17", "content": "@Salamat, can you please explain out what exactly new_merchant transactions are in simple terms please ? I dind't quite understand the way it's mentioned in discussion forum and problem statement . \nAlso, does that mean, if any user has made transactions in the period ( covered by new_merchants) with older merchants it is not listed anywhere ( neither historical_transactions nor new_merchants) ?
", "votes": null, "replies": [ { "id": 459712, "author_name": "salamatburj", "author_url": "", "post_date": "01/22/2019 08:20:19", "content": "Sorry, I can't make it simpler. Try to explore data look at the dates and unique merchats in new merchant transaction. Also, make sure that you read overview and explanations of data.
", "votes": null, "replies": [] } ] }, { "id": 459435, "author_name": "calabist", "author_url": "", "post_date": "01/21/2019 18:16:56", "content": "I think new merchant transactions should help in building features that capture degree of loyalty to the core merchant group. I call merchants in the historic transaction a legacy group and merchants in new transactions a diviant group. From that I have built features that rank high on feature importance scale, but it did not translate to the higher LB score... I hope for the time being :)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "458546": "Why, concatenating all transactions doesn't improve the model, but individually is better?", "458657": "Cuz they are different. Historical transactions are transaction which were made before any discounts(promotion) , while new_merchant transactions are new transactions after discount is given, and only new merchants are included.", "458667": "Salamat, can you please explain out what exactly new_merchant transactions are in simple terms please ? I dind't quite understand the way it's mentioned in discussion forum and problem statement . \nAlso, does that mean, if any user has made transactions in the period ( covered by new_merchants) with older merchants it is not listed anywhere ( neither historical_transactions nor new_merchants) ?", "458855": "May I know how you get this conclusion? TQ", "458975": "i made a model approach according to the fact that new merchants transactions is an evaluation perdiod and still there are incosistencies on the training dataset. For example card_id=C_ID_29d6a47e35, has not transactions at all on the new merchants file and has a loyalty score of 2.584628, which is pretty good score. This card has it's latest transaction really late, on 11/6/2017\n\n2 months now, i try to understand what is the target, but is seems that the competition goal is to find the \"what is the target\" :D \n\nThe opposite happens for card=C_ID_3f7ba15ec2, where has target -17.49263 and on the new merchants file, this card visited 7 different merchants, and on the one had a purchase amount of 1.5, which was the specific merchant's highest transaction amount on the new merchants file(M_ID_f40eec7eb5).... i believe that it is not clear what we are predicting and how the data were simulated. Maybe we are trying to predict noise.... who knows.....", "459435": "I think new merchant transactions should help in building features that capture degree of loyalty to the core merchant group. I call merchants in the historic transaction a legacy group and merchants in new transactions a diviant group. From that I have built features that rank high on feature importance scale, but it did not translate to the higher LB score... I hope for the time being :)", "459707": "skw1990:\nYou can check if you look at new merchant transaction's merchants they will be all unique and almost all of them not present history transactions. There will be some examples where merchants in historical transaction also appear in new merchant transactions but number is very low ~1 (same merchant) for each card_id.", "459710": "Dimosthenis Karaflos\nI am not really sure whether positive target is good or not (maybe in fact smaller is better). My understanding is that we are predicting customer loyality score based on new merchant. However, there are cases that you mentioned. I am not sure how to explain them.", "459712": "Sorry, I can't make it simpler. Try to explore data look at the dates and unique merchats in new merchant transaction. Also, make sure that you read overview and explanations of data.", "459713": "salamat\nIndeed, the data seem irrelevant to the target and the host does not give clear explanations about the target and the normalization process.", "459718": "I think what they are trying to see whether we can predict customer loyalty score for new merchants based on behavior of customer .From historical transactions one might understand about customers preference and behavior. This might also true for new merchant but I think we need it to find some correlations(with historical transactions) which I haven't really tried yet. So, problem is not straightforward we don't really know which features are relevant to customer loyalty for new merchants. I believe problem would be much simpler if we predict customer loyalty score for \"old\" merchant where more information is available(historical transactions).", "459741": "In fact, i tried to read about ELO, and i think that the recommendations are not personalized. I believe that it is like discounts of 20% for example, on a specific merchant. But this discount is available for all the cardholders, so collaboration filtering, FMs,MFdo not seem appropriate methodology if this is the case.", "463249": "Thanks.\n\nBased on my observations, at card level, merchants in old & new transactions are not overlapped; Also, some customers do not have transactions, yet appear to have high score....\nActually I still do not comprehend the target. The loyalty score here is target for merchant or target for customers.....puzzling haha" }, "source": "meta" }