If we partition customers into two groups 'online' and 'stores', it becomes clear that it is more difficult to predict future purchases of offline users than to predict those of online users.
\nIn this notebook, MAP@12 for online users = 0.0287, but for store users = 0.0165.
What could be reasons/causes of this difference and how can we overcome?
", "messages": [ { "id": "1727271", "postDate": "03/17/2022 20:26:51", "content": "If we partition customers into two groups 'online' and 'stores', it becomes clear that it is more difficult to predict future purchases of offline users than to predict those of online users.
\nIn this notebook, MAP@12 for online users = 0.0287, but for store users = 0.0165.
What could be reasons/causes of this difference and how can we overcome?
", "rawMarkdown": "If we partition customers into two groups '**online**' and '**stores**', it becomes clear that it is more difficult to predict future purchases of offline users than to predict those of online users.\nIn this [notebook](https://www.kaggle.com/negoto/h-m-framework-for-partitioned-validation), MAP@12 for online users = 0.0287, but for store users = 0.0165.\n\nWhat could be reasons/causes of this difference and how can we overcome?", "votes": null }, { "id": "1727757", "postDate": "03/18/2022 08:51:09", "content": "Perhaps because the offline shopping experience is much more varied. There are 4850 stores, which may all have different layouts, assortments, stock situations. Compare that to a handful of websites where everybody has roughly the same experience, and probably see the same set of recommendations for a given product at a given timespan
", "rawMarkdown": "Perhaps because the offline shopping experience is much more varied. There are 4850 stores, which may all have different layouts, assortments, stock situations. Compare that to a handful of websites where everybody has roughly the same experience, and probably see the same set of recommendations for a given product at a given timespan", "votes": null }, { "id": "1729399", "postDate": "03/20/2022 03:02:23", "content": "The current high-scoring kernels are in large part based on customers rebuying what they purchased in the past.
\nI believe this is more common for online users - you sign into your account, go to your purchase history, and reorder.
\nAlso - it's been pointed out that some of the repurchases may be from people who mistakenly bought the wrong size, and returned it and then purchased the product again in a different size. Again, this is more likely for online users, who couldn't visually inspect the size before purchasing.
In my opinion, the biggest difference in product choices between online and offline stores are recommendation systems and product names/categories.
", "rawMarkdown": "In my opinion, the biggest difference in product choices between online and offline stores are recommendation systems and product names/categories.", "votes": null }, { "id": "1732986", "postDate": "03/23/2022 23:12:46", "content": "Good sharing
", "rawMarkdown": "Good sharing", "votes": null }, { "id": "1735354", "postDate": "03/26/2022 06:27:34", "content": "Thanks for replies!
\nI am trying to improve my model and each reply has been very helpful.
\nThey've deepened my feelings on the importance of considering the differences between online and in-store transactions.
Great insight! Just want to add one more thing to your list, which is I think each store also makes independent promotion based on different factors.
", "rawMarkdown": "Great insight! Just want to add one more thing to your list, which is I think each store also makes independent promotion based on different factors.", "votes": null }, { "id": "1737253", "postDate": "03/28/2022 09:23:14", "content": "Hi, Ryo!
\nCan u help me to find description of data?
\nI have started competition recently)
Hello @negoto , thank you for your sharing your idea and notebook. My idea about difficulty of prediction for offline purchasing is mainly limited lineup and stock availability at shops. Due to this limitation, customer might buy little bit different article from their standard favorite. ( E.g. customer looking for black shirts but no stock at shop. Then they picked up navy one instead of black.) So I am thinking that for offline it would be good to recommend not only popular article but also similar articles more than online. I am happy if you would share your opinion about that.
", "rawMarkdown": "Hello @negoto , thank you for your sharing your idea and notebook. My idea about difficulty of prediction for offline purchasing is mainly limited lineup and stock availability at shops. Due to this limitation, customer might buy little bit different article from their standard favorite. ( E.g. customer looking for black shirts but no stock at shop. Then they picked up navy one instead of black.) So I am thinking that for offline it would be good to recommend not only popular article but also similar articles more than online. I am happy if you would share your opinion about that.", "votes": null }, { "id": "1737451", "postDate": "03/28/2022 12:48:18", "content": "Probably, this discussion might help you most.
", "rawMarkdown": "Probably, [this discussion](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/307001) might help you most.", "votes": null }, { "id": "1737518", "postDate": "03/28/2022 13:56:20", "content": "Thanks for sharing your idea!
\nI agree that what customers can get at stores depends on the stores' stock. This point might matter.
In my opinion, basically, it is better to choose 'the next most popular' items than 'similar but less popular' articles, for we can choose only 12 items to each customer. However, 'similar to the most popular' may be a good reason to prefer an item over an equally popular one.
", "rawMarkdown": "Thanks for sharing your idea! \nI agree that what customers can get at stores depends on the stores' stock. This point might matter.\n\nIn my opinion, basically, it is better to choose 'the next most popular' items than 'similar but less popular' articles, for we can choose only 12 items to each customer. However, 'similar to the most popular' may be a good reason to prefer an item over an equally popular one.", "votes": null }, { "id": "1737579", "postDate": "03/28/2022 14:42:07", "content": "Hello @negoto , thank you for your prompt feedback. I agree on your idea that it might be good to see ' similar to the most popular '! This would be good hint to think about my next step. Thanks again 🙏
", "rawMarkdown": "Hello @negoto , thank you for your prompt feedback. I agree on your idea that it might be good to see ' similar to the most popular '! This would be good hint to think about my next step. Thanks again 🙏", "votes": null }, { "id": "1738129", "postDate": "03/29/2022 03:58:11", "content": "Thank you for sharing your thought.
\nI checked this.
\nAll transaction data, the online and offline shopping ratio is 70.4% : 29.6%.
\nRebuying transaction data, the ratio is 81.9% : 18.1%.
I would guess that this is mostly driven by a typical online \"phenomenon\": people rebuying the same article in a different size.
", "rawMarkdown": "I would guess that this is mostly driven by a typical online \"phenomenon\": people rebuying the same article in a different size.", "votes": null }, { "id": "1747285", "postDate": "04/06/2022 14:36:43", "content": "Thanks)
\nEstimate my EDA notebook of this competition, maybe u find something interesting
\nhttps://www.kaggle.com/code/matveyspiridonov/h-m-little-eda
Perhaps because the offline shopping experience is much more varied. There are 4850 stores, which may all have different layouts, assortments, stock situations. Compare that to a handful of websites where everybody has roughly the same experience, and probably see the same set of recommendations for a given product at a given timespan
", "votes": null, "replies": [ { "id": 1736782, "author_name": "calmcode", "author_url": "", "post_date": "03/27/2022 18:24:16", "content": "Great insight! Just want to add one more thing to your list, which is I think each store also makes independent promotion based on different factors.
", "votes": null, "replies": [] } ] }, { "id": 1729399, "author_name": "jacob34", "author_url": "", "post_date": "03/20/2022 03:02:23", "content": "The current high-scoring kernels are in large part based on customers rebuying what they purchased in the past.
\nI believe this is more common for online users - you sign into your account, go to your purchase history, and reorder.
\nAlso - it's been pointed out that some of the repurchases may be from people who mistakenly bought the wrong size, and returned it and then purchased the product again in a different size. Again, this is more likely for online users, who couldn't visually inspect the size before purchasing.
Good sharing
", "votes": null, "replies": [] }, { "id": 1738129, "author_name": "ryotak12", "author_url": "", "post_date": "03/29/2022 03:58:11", "content": "Thank you for sharing your thought.
\nI checked this.
\nAll transaction data, the online and offline shopping ratio is 70.4% : 29.6%.
\nRebuying transaction data, the ratio is 81.9% : 18.1%.
In my opinion, the biggest difference in product choices between online and offline stores are recommendation systems and product names/categories.
", "votes": null, "replies": [] }, { "id": 1735354, "author_name": "negoto", "author_url": "", "post_date": "03/26/2022 06:27:34", "content": "Thanks for replies!
\nI am trying to improve my model and each reply has been very helpful.
\nThey've deepened my feelings on the importance of considering the differences between online and in-store transactions.
Hi, Ryo!
\nCan u help me to find description of data?
\nI have started competition recently)
Probably, this discussion might help you most.
", "votes": null, "replies": [] }, { "id": 1747285, "author_name": "matveyspiridonov", "author_url": "", "post_date": "04/06/2022 14:36:43", "content": "Thanks)
\nEstimate my EDA notebook of this competition, maybe u find something interesting
\nhttps://www.kaggle.com/code/matveyspiridonov/h-m-little-eda
Hello @negoto , thank you for your sharing your idea and notebook. My idea about difficulty of prediction for offline purchasing is mainly limited lineup and stock availability at shops. Due to this limitation, customer might buy little bit different article from their standard favorite. ( E.g. customer looking for black shirts but no stock at shop. Then they picked up navy one instead of black.) So I am thinking that for offline it would be good to recommend not only popular article but also similar articles more than online. I am happy if you would share your opinion about that.
", "votes": null, "replies": [ { "id": 1737518, "author_name": "negoto", "author_url": "", "post_date": "03/28/2022 13:56:20", "content": "Thanks for sharing your idea!
\nI agree that what customers can get at stores depends on the stores' stock. This point might matter.
In my opinion, basically, it is better to choose 'the next most popular' items than 'similar but less popular' articles, for we can choose only 12 items to each customer. However, 'similar to the most popular' may be a good reason to prefer an item over an equally popular one.
", "votes": null, "replies": [] }, { "id": 1737579, "author_name": "hechtjp", "author_url": "", "post_date": "03/28/2022 14:42:07", "content": "Hello @negoto , thank you for your prompt feedback. I agree on your idea that it might be good to see ' similar to the most popular '! This would be good hint to think about my next step. Thanks again 🙏
", "votes": null, "replies": [] } ] }, { "id": 1738402, "author_name": "tbierhance", "author_url": "", "post_date": "03/29/2022 08:55:06", "content": "I would guess that this is mostly driven by a typical online \"phenomenon\": people rebuying the same article in a different size.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1727271": "If we partition customers into two groups '**online**' and '**stores**', it becomes clear that it is more difficult to predict future purchases of offline users than to predict those of online users.\nIn this [notebook](https://www.kaggle.com/negoto/h-m-framework-for-partitioned-validation), MAP@12 for online users = 0.0287, but for store users = 0.0165.\n\nWhat could be reasons/causes of this difference and how can we overcome?", "1727757": "Perhaps because the offline shopping experience is much more varied. There are 4850 stores, which may all have different layouts, assortments, stock situations. Compare that to a handful of websites where everybody has roughly the same experience, and probably see the same set of recommendations for a given product at a given timespan", "1729399": "The current high-scoring kernels are in large part based on customers rebuying what they purchased in the past.\n\nI believe this is more common for online users - you sign into your account, go to your purchase history, and reorder.\nAlso - it's been pointed out that some of the repurchases may be from people who mistakenly bought the wrong size, and returned it and then purchased the product again in a different size. Again, this is more likely for online users, who couldn't visually inspect the size before purchasing.", "1732445": "In my opinion, the biggest difference in product choices between online and offline stores are recommendation systems and product names/categories.", "1732986": "Good sharing", "1735354": "Thanks for replies!\nI am trying to improve my model and each reply has been very helpful.\nThey've deepened my feelings on the importance of considering the differences between online and in-store transactions.", "1736782": "Great insight! Just want to add one more thing to your list, which is I think each store also makes independent promotion based on different factors.", "1737253": "Hi, Ryo!\nCan u help me to find description of data? \nI have started competition recently)", "1737273": "Hello @negoto , thank you for your sharing your idea and notebook. My idea about difficulty of prediction for offline purchasing is mainly limited lineup and stock availability at shops. Due to this limitation, customer might buy little bit different article from their standard favorite. ( E.g. customer looking for black shirts but no stock at shop. Then they picked up navy one instead of black.) So I am thinking that for offline it would be good to recommend not only popular article but also similar articles more than online. I am happy if you would share your opinion about that.", "1737451": "Probably, [this discussion](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/307001) might help you most.", "1737518": "Thanks for sharing your idea! \nI agree that what customers can get at stores depends on the stores' stock. This point might matter.\n\nIn my opinion, basically, it is better to choose 'the next most popular' items than 'similar but less popular' articles, for we can choose only 12 items to each customer. However, 'similar to the most popular' may be a good reason to prefer an item over an equally popular one.", "1737579": "Hello @negoto , thank you for your prompt feedback. I agree on your idea that it might be good to see ' similar to the most popular '! This would be good hint to think about my next step. Thanks again 🙏", "1738129": "Thank you for sharing your thought.\nI checked this.\nAll transaction data, the online and offline shopping ratio is 70.4% : 29.6%.\nRebuying transaction data, the ratio is 81.9% : 18.1%.", "1738402": "I would guess that this is mostly driven by a typical online \"phenomenon\": people rebuying the same article in a different size.", "1747285": "Thanks)\nEstimate my EDA notebook of this competition, maybe u find something interesting\nhttps://www.kaggle.com/code/matveyspiridonov/h-m-little-eda" }, "source": "meta" }