Imagine you work in a clothing store and your best friend walks in.
\nYou know everything there is to know about her – what type of clothing she wears (and her children wear), what her taste is in brand, style and color, what her financial situation and preferences are, the age of her children – even her recent purchases and what’s currently missing from her wardrobe!
\nYour boss asks you to go over and give her some recommendations. Given your knowledge of what the store has to offer, what your friend's preferences are, and what people with similar preferences to your friend have been buying recently, it’s likely she’d seriously consider many of the recommendations you’d make to her. Not only that, but it’s likely that you could come up with a broad range of products that would appeal to her, not necessarily a specific one or two.
\nBut now let’s consider a different scenario. It’s a busy day in the store, and when your friend comes in, you’re working the cash register, and don’t even have time to even send a hand-wave her way. The question crosses your mind – what will your friend show up to checkout with?
\nNow, that’s a different question. Of the 100s of recommendations you might have made, it’s likely that she won’t even see most of them. She may not even visit their section of the store. What she sees will depend on what she came to the store for, which sections she’s visited in the past, which items are prominently displayed, and which are part of a prominent display etc.
\nIt’s also a much harder question – it’s not enough to come up with any of a few hundred items she’d be interested in, rather you need to guess the actual few items she’s going to buy.
\nThis competition is presented as being about recommendations. But looking at what we need to predict and how it’s evaluated, it seems to be a different kind of challenge, and a broader and harder one at that.
\nI think we need to keep that in mind before coding up some classic recommendation models.
", "messages": [ { "id": "1683765", "postDate": "02/10/2022 03:27:22", "content": "Imagine you work in a clothing store and your best friend walks in.
\nYou know everything there is to know about her – what type of clothing she wears (and her children wear), what her taste is in brand, style and color, what her financial situation and preferences are, the age of her children – even her recent purchases and what’s currently missing from her wardrobe!
\nYour boss asks you to go over and give her some recommendations. Given your knowledge of what the store has to offer, what your friend's preferences are, and what people with similar preferences to your friend have been buying recently, it’s likely she’d seriously consider many of the recommendations you’d make to her. Not only that, but it’s likely that you could come up with a broad range of products that would appeal to her, not necessarily a specific one or two.
\nBut now let’s consider a different scenario. It’s a busy day in the store, and when your friend comes in, you’re working the cash register, and don’t even have time to even send a hand-wave her way. The question crosses your mind – what will your friend show up to checkout with?
\nNow, that’s a different question. Of the 100s of recommendations you might have made, it’s likely that she won’t even see most of them. She may not even visit their section of the store. What she sees will depend on what she came to the store for, which sections she’s visited in the past, which items are prominently displayed, and which are part of a prominent display etc.
\nIt’s also a much harder question – it’s not enough to come up with any of a few hundred items she’d be interested in, rather you need to guess the actual few items she’s going to buy.
\nThis competition is presented as being about recommendations. But looking at what we need to predict and how it’s evaluated, it seems to be a different kind of challenge, and a broader and harder one at that.
\nI think we need to keep that in mind before coding up some classic recommendation models.
", "rawMarkdown": "Imagine you work in a clothing store and your best friend walks in.\n\nYou know everything there is to know about her – what type of clothing she wears (and her children wear), what her taste is in brand, style and color, what her financial situation and preferences are, the age of her children – even her recent purchases and what’s currently missing from her wardrobe! \n\nYour boss asks you to go over and give her some recommendations. Given your knowledge of what the store has to offer, what your friend's preferences are, and what people with similar preferences to your friend have been buying recently, it’s likely she’d seriously consider many of the recommendations you’d make to her. Not only that, but it’s likely that you could come up with a broad range of products that would appeal to her, not necessarily a specific one or two.\n\nBut now let’s consider a different scenario. It’s a busy day in the store, and when your friend comes in, you’re working the cash register, and don’t even have time to even send a hand-wave her way. The question crosses your mind – what will your friend show up to checkout with?\n\nNow, that’s a different question. Of the 100s of recommendations you might have made, it’s likely that she won’t even see most of them. She may not even visit their section of the store. What she sees will depend on what she came to the store for, which sections she’s visited in the past, which items are prominently displayed, and which are part of a prominent display etc.\n\nIt’s also a much harder question – it’s not enough to come up with any of a few hundred items she’d be interested in, rather you need to guess the actual few items she’s going to buy.\n\nThis competition is presented as being about **recommendations**. But looking at what we need to predict and how it’s evaluated, it seems to be a different kind of challenge, and a broader and harder one at that.\n\nI think we need to keep that in mind before coding up some classic recommendation models.", "votes": null }, { "id": "1684954", "postDate": "02/10/2022 22:11:16", "content": "It seems to me like the winning solution will be the model that predicts what the customers will buy next rather than just what they would probably rate the highest. That's influence by season, price changes, and other trends that I don't think would be easily captured by a traditional recommendation system.
", "rawMarkdown": "It seems to me like the winning solution will be the model that predicts what the customers will buy next rather than just what they would probably rate the highest. That's influence by season, price changes, and other trends that I don't think would be easily captured by a traditional recommendation system.", "votes": null }, { "id": "1685439", "postDate": "02/11/2022 09:19:23", "content": "Totally agree with your opinion. I think this task is the consumer behavior prediction, not recommendation.
", "rawMarkdown": "Totally agree with your opinion. I think this task is **the consumer behavior prediction**, not recommendation.", "votes": null }, { "id": "1686279", "postDate": "02/11/2022 22:34:52", "content": "This is what they say in introduction
\n\"In this competition, H&M Group invites you to develop product recommendations based on data from previous transactions, as well as from customer and product meta data. The available meta data spans from simple data, such as garment type and customer age, to text data from product descriptions, to image data from garment images.
\nThere are no preconceptions on what information that may be useful – that is for you to find out. If you want to investigate a categorical data type algorithm, or dive into NLP and image processing deep learning, that is up to you.\"
\nI can understand that for example, the product descriptions, may offer clues to the preferences of a customer and help in making recommendations. For images though, would it make sense to categorize different images and build preferences based on that?
", "rawMarkdown": "This is what they say in introduction\n\n\"In this competition, H&M Group invites you to develop product recommendations based on data from previous transactions, as well as from customer and product meta data. The available meta data spans from simple data, such as garment type and customer age, to text data from product descriptions, to image data from garment images.\n\nThere are no preconceptions on what information that may be useful – that is for you to find out. If you want to investigate a categorical data type algorithm, or dive into NLP and image processing deep learning, that is up to you.\"\n\nI can understand that for example, the product descriptions, may offer clues to the preferences of a customer and help in making recommendations. For images though, would it make sense to categorize different images and build preferences based on that?", "votes": null }, { "id": "1692344", "postDate": "02/16/2022 01:37:27", "content": "Perhaps there's likeness between garments that an image processing model will detect but that simply isn't there in the descriptions?
", "rawMarkdown": "Perhaps there's likeness between garments that an image processing model will detect but that simply isn't there in the descriptions?", "votes": null }, { "id": "1692355", "postDate": "02/16/2022 01:57:46", "content": "Some observations:
\n1 Some people buy \"the same\" t-shirt and pair of jeans for decades. Some people won't even buy socks of the same colour twice. How can that nuance be detected in the data and modeled?
\n2 Some people shop whole outfits at once. Some people supplement what they already have in order to be able to combine new outfits. Some people don't care about outfits at all. How can that nuance be detected in the data and modeled?
\n3 The market segment H&M operates in has customers of many different approaches to clothing, style and fashion. There's Fast Fashionistas that buys clothes on a weekly basis. There's the more Style conscious tribe, that don't do their main shopping at fast fashion retailers, but will supplement with basic garments or the odd novelty item. There's some that don't really care but appreciate the value at an affordable price. How can we sort/classify customers into different fashion tribes by means of clustering and the like?
\nI'm a newbie when it comes to ML, but I do know a thing or two about the clothing retail market and style.
\nWill domain knowledge be a major factor in winning solutions?
", "rawMarkdown": "Some observations:\n\n**1** Some people buy \"the same\" t-shirt and pair of jeans for decades. Some people won't even buy socks of the same colour twice. How can that nuance be detected in the data and modeled?\n\n**2** Some people shop whole outfits at once. Some people supplement what they already have in order to be able to combine new outfits. Some people don't care about outfits at all. How can that nuance be detected in the data and modeled?\n\n**3** The market segment H&M operates in has customers of many different approaches to clothing, style and fashion. There's *Fast Fashionistas* that buys clothes on a weekly basis. There's the more *Style conscious* tribe, that don't do their main shopping at fast fashion retailers, but will supplement with basic garments or the odd novelty item. There's some that don't really care but appreciate the value at an affordable price. **How can we sort/classify customers into different *fashion tribes* by means of clustering and the like?**\n\nI'm a newbie when it comes to ML, but I do know a thing or two about the clothing retail market and style.\n\nWill domain knowledge be a major factor in winning solutions?", "votes": null }, { "id": "1692430", "postDate": "02/16/2022 03:23:20", "content": "\n\nI'm a newbie when it comes to ML, but I do know a thing or two about the clothing retail market and style.
\nWill domain knowledge be a major factor in winning solutions?
\n
I don't know - in most competitions, I haven't noticed that it was.
\nAs far as if it can help you - certainly. Tabular data is a lot about feature engineering - calculating new features from the ones provided - and domain knowledge gives you insight into what features will be helpful.
\nI would advise to first look around at how it's being approached from a ML perspective in the shared notebooks (and there will be more than one approach), and then see if your domain knowledge is applicable within that framework.
", "rawMarkdown": "> I'm a newbie when it comes to ML, but I do know a thing or two about the clothing retail market and style.\n\n> Will domain knowledge be a major factor in winning solutions?\n\nI don't know - in most competitions, I haven't noticed that it was.\n\nAs far as if it can help you - certainly. Tabular data is a lot about feature engineering - calculating new features from the ones provided - and domain knowledge gives you insight into what features will be helpful.\n\nI would advise to first look around at how it's being approached from a ML perspective in the shared notebooks (and there will be more than one approach), and then see if your domain knowledge is applicable within that framework.", "votes": null } ], "comments": [ { "id": 1684954, "author_name": "jacobwregan", "author_url": "", "post_date": "02/10/2022 22:11:16", "content": "It seems to me like the winning solution will be the model that predicts what the customers will buy next rather than just what they would probably rate the highest. That's influence by season, price changes, and other trends that I don't think would be easily captured by a traditional recommendation system.
", "votes": null, "replies": [] }, { "id": 1685439, "author_name": "tj0612", "author_url": "", "post_date": "02/11/2022 09:19:23", "content": "Totally agree with your opinion. I think this task is the consumer behavior prediction, not recommendation.
", "votes": null, "replies": [] }, { "id": 1686279, "author_name": "atulverma", "author_url": "", "post_date": "02/11/2022 22:34:52", "content": "This is what they say in introduction
\n\"In this competition, H&M Group invites you to develop product recommendations based on data from previous transactions, as well as from customer and product meta data. The available meta data spans from simple data, such as garment type and customer age, to text data from product descriptions, to image data from garment images.
\nThere are no preconceptions on what information that may be useful – that is for you to find out. If you want to investigate a categorical data type algorithm, or dive into NLP and image processing deep learning, that is up to you.\"
\nI can understand that for example, the product descriptions, may offer clues to the preferences of a customer and help in making recommendations. For images though, would it make sense to categorize different images and build preferences based on that?
", "votes": null, "replies": [ { "id": 1692344, "author_name": "nimadjoharitaimouri", "author_url": "", "post_date": "02/16/2022 01:37:27", "content": "Perhaps there's likeness between garments that an image processing model will detect but that simply isn't there in the descriptions?
", "votes": null, "replies": [] } ] }, { "id": 1692355, "author_name": "nimadjoharitaimouri", "author_url": "", "post_date": "02/16/2022 01:57:46", "content": "Some observations:
\n1 Some people buy \"the same\" t-shirt and pair of jeans for decades. Some people won't even buy socks of the same colour twice. How can that nuance be detected in the data and modeled?
\n2 Some people shop whole outfits at once. Some people supplement what they already have in order to be able to combine new outfits. Some people don't care about outfits at all. How can that nuance be detected in the data and modeled?
\n3 The market segment H&M operates in has customers of many different approaches to clothing, style and fashion. There's Fast Fashionistas that buys clothes on a weekly basis. There's the more Style conscious tribe, that don't do their main shopping at fast fashion retailers, but will supplement with basic garments or the odd novelty item. There's some that don't really care but appreciate the value at an affordable price. How can we sort/classify customers into different fashion tribes by means of clustering and the like?
\nI'm a newbie when it comes to ML, but I do know a thing or two about the clothing retail market and style.
\nWill domain knowledge be a major factor in winning solutions?
", "votes": null, "replies": [ { "id": 1692430, "author_name": "jacob34", "author_url": "", "post_date": "02/16/2022 03:23:20", "content": "\n\nI'm a newbie when it comes to ML, but I do know a thing or two about the clothing retail market and style.
\nWill domain knowledge be a major factor in winning solutions?
\n
I don't know - in most competitions, I haven't noticed that it was.
\nAs far as if it can help you - certainly. Tabular data is a lot about feature engineering - calculating new features from the ones provided - and domain knowledge gives you insight into what features will be helpful.
\nI would advise to first look around at how it's being approached from a ML perspective in the shared notebooks (and there will be more than one approach), and then see if your domain knowledge is applicable within that framework.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1683765": "Imagine you work in a clothing store and your best friend walks in.\n\nYou know everything there is to know about her – what type of clothing she wears (and her children wear), what her taste is in brand, style and color, what her financial situation and preferences are, the age of her children – even her recent purchases and what’s currently missing from her wardrobe! \n\nYour boss asks you to go over and give her some recommendations. Given your knowledge of what the store has to offer, what your friend's preferences are, and what people with similar preferences to your friend have been buying recently, it’s likely she’d seriously consider many of the recommendations you’d make to her. Not only that, but it’s likely that you could come up with a broad range of products that would appeal to her, not necessarily a specific one or two.\n\nBut now let’s consider a different scenario. It’s a busy day in the store, and when your friend comes in, you’re working the cash register, and don’t even have time to even send a hand-wave her way. The question crosses your mind – what will your friend show up to checkout with?\n\nNow, that’s a different question. Of the 100s of recommendations you might have made, it’s likely that she won’t even see most of them. She may not even visit their section of the store. What she sees will depend on what she came to the store for, which sections she’s visited in the past, which items are prominently displayed, and which are part of a prominent display etc.\n\nIt’s also a much harder question – it’s not enough to come up with any of a few hundred items she’d be interested in, rather you need to guess the actual few items she’s going to buy.\n\nThis competition is presented as being about **recommendations**. But looking at what we need to predict and how it’s evaluated, it seems to be a different kind of challenge, and a broader and harder one at that.\n\nI think we need to keep that in mind before coding up some classic recommendation models.", "1684954": "It seems to me like the winning solution will be the model that predicts what the customers will buy next rather than just what they would probably rate the highest. That's influence by season, price changes, and other trends that I don't think would be easily captured by a traditional recommendation system.", "1685439": "Totally agree with your opinion. I think this task is **the consumer behavior prediction**, not recommendation.", "1686279": "This is what they say in introduction\n\n\"In this competition, H&M Group invites you to develop product recommendations based on data from previous transactions, as well as from customer and product meta data. The available meta data spans from simple data, such as garment type and customer age, to text data from product descriptions, to image data from garment images.\n\nThere are no preconceptions on what information that may be useful – that is for you to find out. If you want to investigate a categorical data type algorithm, or dive into NLP and image processing deep learning, that is up to you.\"\n\nI can understand that for example, the product descriptions, may offer clues to the preferences of a customer and help in making recommendations. For images though, would it make sense to categorize different images and build preferences based on that?", "1692344": "Perhaps there's likeness between garments that an image processing model will detect but that simply isn't there in the descriptions?", "1692355": "Some observations:\n\n**1** Some people buy \"the same\" t-shirt and pair of jeans for decades. Some people won't even buy socks of the same colour twice. How can that nuance be detected in the data and modeled?\n\n**2** Some people shop whole outfits at once. Some people supplement what they already have in order to be able to combine new outfits. Some people don't care about outfits at all. How can that nuance be detected in the data and modeled?\n\n**3** The market segment H&M operates in has customers of many different approaches to clothing, style and fashion. There's *Fast Fashionistas* that buys clothes on a weekly basis. There's the more *Style conscious* tribe, that don't do their main shopping at fast fashion retailers, but will supplement with basic garments or the odd novelty item. There's some that don't really care but appreciate the value at an affordable price. **How can we sort/classify customers into different *fashion tribes* by means of clustering and the like?**\n\nI'm a newbie when it comes to ML, but I do know a thing or two about the clothing retail market and style.\n\nWill domain knowledge be a major factor in winning solutions?", "1692430": "> I'm a newbie when it comes to ML, but I do know a thing or two about the clothing retail market and style.\n\n> Will domain knowledge be a major factor in winning solutions?\n\nI don't know - in most competitions, I haven't noticed that it was.\n\nAs far as if it can help you - certainly. Tabular data is a lot about feature engineering - calculating new features from the ones provided - and domain knowledge gives you insight into what features will be helpful.\n\nI would advise to first look around at how it's being approached from a ML perspective in the shared notebooks (and there will be more than one approach), and then see if your domain knowledge is applicable within that framework." }, "source": "meta" }