I was playing around with the current best public notebook (time is our best friend V2), and basically the idea is to recommend the 12 most bought products by a customer during the past 1, 2 or 3 weeks. and if the customer did not make a purchase in the last 3 weeks we would recommend the most popular articles during the last week.
\nThis is a quite good heuristic and gives 0.02 on the public LB (even with some variations, the score is still 0.02).
\nWhat I don't understand however is that why recommending the same products already bought giving a good result ? If a customer already bought a product in the last 3, 2 or one week he would probably not buy it again, right ?
\nCould anyone clarify this point please ?
", "messages": [ { "id": "1698387", "postDate": "02/20/2022 10:59:55", "content": "I was playing around with the current best public notebook (time is our best friend V2), and basically the idea is to recommend the 12 most bought products by a customer during the past 1, 2 or 3 weeks. and if the customer did not make a purchase in the last 3 weeks we would recommend the most popular articles during the last week.
\nThis is a quite good heuristic and gives 0.02 on the public LB (even with some variations, the score is still 0.02).
\nWhat I don't understand however is that why recommending the same products already bought giving a good result ? If a customer already bought a product in the last 3, 2 or one week he would probably not buy it again, right ?
\nCould anyone clarify this point please ?
", "rawMarkdown": "I was playing around with the current best public notebook (time is our best friend V2), and basically the idea is to recommend the 12 most bought products by a customer during the past 1, 2 or 3 weeks. and if the customer did not make a purchase in the last 3 weeks we would recommend the most popular articles during the last week.\n\nThis is a quite good heuristic and gives 0.02 on the public LB (even with some variations, the score is still 0.02).\n\nWhat I don't understand however is that why recommending the same products already bought giving a good result ? If a customer already bought a product in the last 3, 2 or one week he would probably not buy it again, right ?\n\nCould anyone clarify this point please ?", "votes": null }, { "id": "1698395", "postDate": "02/20/2022 11:06:50", "content": "@souamesannis I might be wrong but many people like owning multiple articles of the same clothing, H&M's target audience might be in that set
", "rawMarkdown": "souamesannis I might be wrong but many people like owning multiple articles of the same clothing, H&M's target audience might be in that set", "votes": null }, { "id": "1698662", "postDate": "02/20/2022 15:05:00", "content": "you could look at what articles are actually being bought more than once. My guess is that whereas a night dress is probably a single purchase, other articles can be socks, stockings etc where buying multiples might make sense.
", "rawMarkdown": "you could look at what articles are actually being bought more than once. My guess is that whereas a night dress is probably a single purchase, other articles can be socks, stockings etc where buying multiples might make sense.", "votes": null }, { "id": "1698665", "postDate": "02/20/2022 15:07:42", "content": "I think it is relatively uncommon for it to happen, but it sometimes will - customers realize they need more of a certain item they bought, and they already bought it, so they know that they like it and where to find it.
\n.02 is not that \"good\" a score - it means we're only getting a small percentage of the purchases.
\nIt's just that these predictions are low-hanging fruit - it's not easy to figure out which product out of 100,000 items a customer will buy, but it's easy to find items they bought already.
", "rawMarkdown": "I think it is relatively uncommon for it to happen, but it sometimes will - customers realize they need more of a certain item they bought, and they already bought it, so they know that they like it and where to find it.\n\n.02 is not that \"good\" a score - it means we're only getting a small percentage of the purchases.\n\nIt's just that these predictions are low-hanging fruit - it's not easy to figure out which product out of 100,000 items a customer will buy, but it's easy to find items they bought already.", "votes": null }, { "id": "1699085", "postDate": "02/20/2022 22:49:09", "content": "I think it's due to returns and rebuys due to size mismatches.
", "rawMarkdown": "I think it's due to returns and rebuys due to size mismatches.", "votes": null }, { "id": "1701607", "postDate": "02/22/2022 23:36:11", "content": "Agree. So if the this type of recommendation is further nuanced based on the product type, it may get a better score.
", "rawMarkdown": "Agree. So if the this type of recommendation is further nuanced based on the product type, it may get a better score.", "votes": null }, { "id": "1701611", "postDate": "02/22/2022 23:39:09", "content": "Leaderboard is .031 at the moment. Not a huge difference. It will be interesting to see what the winning score would be.
", "rawMarkdown": "Leaderboard is .031 at the moment. Not a huge difference. It will be interesting to see what the winning score would be.", "votes": null }, { "id": "1701612", "postDate": "02/22/2022 23:40:45", "content": "It should be a small percentage. Also, in real world, it would not actually be a new sale and recommendation would also not be needed for that.
", "rawMarkdown": "It should be a small percentage. Also, in real world, it would not actually be a new sale and recommendation would also not be needed for that.", "votes": null }, { "id": "1701617", "postDate": "02/22/2022 23:43:34", "content": "Then the whole overall recommendation should be very simple..I would think. Categorize customers and their purchases and recommend the same product over an over again. For one, it will have to end sometime. Second, it would not be much of a data science / ML task.
", "rawMarkdown": "Then the whole overall recommendation should be very simple..I would think. Categorize customers and their purchases and recommend the same product over an over again. For one, it will have to end sometime. Second, it would not be much of a data science / ML task.", "votes": null } ], "comments": [ { "id": 1698395, "author_name": "init27", "author_url": "", "post_date": "02/20/2022 11:06:50", "content": "@souamesannis I might be wrong but many people like owning multiple articles of the same clothing, H&M's target audience might be in that set
", "votes": null, "replies": [ { "id": 1701617, "author_name": "atulverma", "author_url": "", "post_date": "02/22/2022 23:43:34", "content": "Then the whole overall recommendation should be very simple..I would think. Categorize customers and their purchases and recommend the same product over an over again. For one, it will have to end sometime. Second, it would not be much of a data science / ML task.
", "votes": null, "replies": [] } ] }, { "id": 1698662, "author_name": "darkbitur", "author_url": "", "post_date": "02/20/2022 15:05:00", "content": "you could look at what articles are actually being bought more than once. My guess is that whereas a night dress is probably a single purchase, other articles can be socks, stockings etc where buying multiples might make sense.
", "votes": null, "replies": [ { "id": 1701607, "author_name": "atulverma", "author_url": "", "post_date": "02/22/2022 23:36:11", "content": "Agree. So if the this type of recommendation is further nuanced based on the product type, it may get a better score.
", "votes": null, "replies": [] } ] }, { "id": 1698665, "author_name": "jacob34", "author_url": "", "post_date": "02/20/2022 15:07:42", "content": "I think it is relatively uncommon for it to happen, but it sometimes will - customers realize they need more of a certain item they bought, and they already bought it, so they know that they like it and where to find it.
\n.02 is not that \"good\" a score - it means we're only getting a small percentage of the purchases.
\nIt's just that these predictions are low-hanging fruit - it's not easy to figure out which product out of 100,000 items a customer will buy, but it's easy to find items they bought already.
", "votes": null, "replies": [ { "id": 1701611, "author_name": "atulverma", "author_url": "", "post_date": "02/22/2022 23:39:09", "content": "Leaderboard is .031 at the moment. Not a huge difference. It will be interesting to see what the winning score would be.
", "votes": null, "replies": [] } ] }, { "id": 1699085, "author_name": "mayukh18", "author_url": "", "post_date": "02/20/2022 22:49:09", "content": "I think it's due to returns and rebuys due to size mismatches.
", "votes": null, "replies": [ { "id": 1701612, "author_name": "atulverma", "author_url": "", "post_date": "02/22/2022 23:40:45", "content": "It should be a small percentage. Also, in real world, it would not actually be a new sale and recommendation would also not be needed for that.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1698387": "I was playing around with the current best public notebook (time is our best friend V2), and basically the idea is to recommend the 12 most bought products by a customer during the past 1, 2 or 3 weeks. and if the customer did not make a purchase in the last 3 weeks we would recommend the most popular articles during the last week.\n\nThis is a quite good heuristic and gives 0.02 on the public LB (even with some variations, the score is still 0.02).\n\nWhat I don't understand however is that why recommending the same products already bought giving a good result ? If a customer already bought a product in the last 3, 2 or one week he would probably not buy it again, right ?\n\nCould anyone clarify this point please ?", "1698395": "souamesannis I might be wrong but many people like owning multiple articles of the same clothing, H&M's target audience might be in that set", "1698662": "you could look at what articles are actually being bought more than once. My guess is that whereas a night dress is probably a single purchase, other articles can be socks, stockings etc where buying multiples might make sense.", "1698665": "I think it is relatively uncommon for it to happen, but it sometimes will - customers realize they need more of a certain item they bought, and they already bought it, so they know that they like it and where to find it.\n\n.02 is not that \"good\" a score - it means we're only getting a small percentage of the purchases.\n\nIt's just that these predictions are low-hanging fruit - it's not easy to figure out which product out of 100,000 items a customer will buy, but it's easy to find items they bought already.", "1699085": "I think it's due to returns and rebuys due to size mismatches.", "1701607": "Agree. So if the this type of recommendation is further nuanced based on the product type, it may get a better score.", "1701611": "Leaderboard is .031 at the moment. Not a huge difference. It will be interesting to see what the winning score would be.", "1701612": "It should be a small percentage. Also, in real world, it would not actually be a new sale and recommendation would also not be needed for that.", "1701617": "Then the whole overall recommendation should be very simple..I would think. Categorize customers and their purchases and recommend the same product over an over again. For one, it will have to end sometime. Second, it would not be much of a data science / ML task." }, "source": "meta" }