{ "id": 324486, "title": "Summarize top-10 place solutions", "url": "/competitions/h-and-m-personalized-fashion-recommendations/discussion/324486", "author_name": "", "post_date": "2022-05-12T01:52:11.992436100Z", "votes": 52, "comment_count": 8, "views": 0, "content": "

Thanks to the organizers for this interesting competition.
\nI read the following articles which are from 1st to 10-th place solutions.
\nThese are very interesting so I summarize the following method to one table.

\n

Links

\n
    \n
  1. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324070
    2. \n
  2. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324197
    3. \n
  3. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324129
    4. \n
  4. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324094
    5. \n
  5. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324098
    6. \n
  6. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324075
    7. \n
  7. -
    8. \n
  8. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324185
    9. \n
  9. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324127
    10. \n
  10. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324223
    11. \n
\n

Table

\n

\"top10-table.png\"

\n

I can't find the articles from 7-th place so I don't fill the 7-th place in the table.

\n

Ranking

\n

Ranking strategies are similar among the top-10 teams.
\nSurprisingly, it seem all teams use LightGBM for the ranking.
\nAt 4-th place, DCN was used in addition to GBDT but it is said that \"The score of LightGBM is much better\".
\nCatBoost was also used by several teams.

\n

For training period, it seems that it's popular to use 4 - 9 weeks and last week was used for the validation.
\nNumber of features are only written in some articles, but about 300 features are added to the models.

\n

Regarding the ensembles, GBDT was commonly used and MLP or different linear model are not used frequently.

\n

Retrieval/Recall

\n

There are a variety of Retrieval/Recall strategies comparing to the ranking.
\nAccording to the 1st place article, it is said that \"the candidate generation strategy is the key to beyond limitation of accuracy, good feature engineering or modeling could be close to the limitation\".

\n

I think this sentence represents this competition and it's very interesting.
\nAccording to the 5-th place article, its is said \"we spend most of the time to develop recall strategies\".

\n

As already mentioned, there are a variety of retrieval method.
\nTherefore, it's difficult to summarize in one table.
\nFor more details, please see the individual great article.

\n

Regarding the retrieval method, the following method are commonly used.

\n\n

These methods had been mentioned in Discussion/Note during the competition.
\nTherefore, I guess a lot of teams used these solutions to their retrieval strategies.

\n

In addition to these method, I guess some ingenuity are needed to get gold medal.

\n

Number of candidate are 100-1000 in most case.

\n

Summary

\n

I summarize from the 1st to 10-th solutions which are informative and exciting.
\nI really appreciate to the author of these articles.

", "messages": [ { "id": "1785286", "postDate": "05/12/2022 01:52:11", "content": "

Thanks to the organizers for this interesting competition.
\nI read the following articles which are from 1st to 10-th place solutions.
\nThese are very interesting so I summarize the following method to one table.

\n

Links

\n
    \n
  1. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324070
    2. \n
  2. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324197
    3. \n
  3. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324129
    4. \n
  4. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324094
    5. \n
  5. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324098
    6. \n
  6. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324075
    7. \n
  7. -
    8. \n
  8. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324185
    9. \n
  9. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324127
    10. \n
  10. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324223
    11. \n
\n

Table

\n

\"top10-table.png\"

\n

I can't find the articles from 7-th place so I don't fill the 7-th place in the table.

\n

Ranking

\n

Ranking strategies are similar among the top-10 teams.
\nSurprisingly, it seem all teams use LightGBM for the ranking.
\nAt 4-th place, DCN was used in addition to GBDT but it is said that \"The score of LightGBM is much better\".
\nCatBoost was also used by several teams.

\n

For training period, it seems that it's popular to use 4 - 9 weeks and last week was used for the validation.
\nNumber of features are only written in some articles, but about 300 features are added to the models.

\n

Regarding the ensembles, GBDT was commonly used and MLP or different linear model are not used frequently.

\n

Retrieval/Recall

\n

There are a variety of Retrieval/Recall strategies comparing to the ranking.
\nAccording to the 1st place article, it is said that \"the candidate generation strategy is the key to beyond limitation of accuracy, good feature engineering or modeling could be close to the limitation\".

\n

I think this sentence represents this competition and it's very interesting.
\nAccording to the 5-th place article, its is said \"we spend most of the time to develop recall strategies\".

\n

As already mentioned, there are a variety of retrieval method.
\nTherefore, it's difficult to summarize in one table.
\nFor more details, please see the individual great article.

\n

Regarding the retrieval method, the following method are commonly used.

\n\n

These methods had been mentioned in Discussion/Note during the competition.
\nTherefore, I guess a lot of teams used these solutions to their retrieval strategies.

\n

In addition to these method, I guess some ingenuity are needed to get gold medal.

\n

Number of candidate are 100-1000 in most case.

\n

Summary

\n

I summarize from the 1st to 10-th solutions which are informative and exciting.
\nI really appreciate to the author of these articles.

", "rawMarkdown": "Thanks to the organizers for this interesting competition.\nI read the following articles which are from 1st to 10-th place solutions.\nThese are very interesting so I summarize the following method to one table.\n\n## Links\n\n1. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324070\n2. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324197\n3. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324129\n4. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324094\n5. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324098\n6. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324075\n7. -\n8. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324185\n9. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324127\n10. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324223\n\n## Table\n\n[![top10-table.png](https://i.postimg.cc/qMLsH2kH/top10-table.png)](https://postimg.cc/RNW6c66g)\n\n\nI can't find the articles from 7-th place so I don't fill the 7-th place in the table.\n\n## Ranking\nRanking strategies are similar among the top-10 teams.\nSurprisingly, it seem all teams use LightGBM for the ranking. \nAt 4-th place, DCN was used in addition to GBDT but it is said that \"The score of LightGBM is much better\".\nCatBoost was also used by several teams.\n\nFor training period, it seems that it's popular to use 4 - 9 weeks and last week was used for the validation.\nNumber of features are only written in some articles, but about 300 features are added to the models.\n\nRegarding the ensembles, GBDT was commonly used and MLP or different linear model are not used frequently.\n\n\n## Retrieval/Recall\n\nThere are a variety of Retrieval/Recall strategies comparing to the ranking.\nAccording to the 1st place article, it is said that \"the candidate generation strategy is the key to beyond limitation of accuracy, good feature engineering or modeling could be close to the limitation\".\n\nI think this sentence represents this competition and it's very interesting.\nAccording to the 5-th place article, its is said \"we spend most of the time to develop recall strategies\".\n\nAs already mentioned, there are a variety of retrieval method.\nTherefore, it's difficult to summarize in one table.\nFor more details, please see the individual great article.\n\nRegarding the retrieval method, the following method are commonly used.\n- repurchase\n- pupular item\n- item collaborative filterling\n\nThese methods had been mentioned in Discussion/Note during the competition.\nTherefore, I guess a lot of teams used these solutions to their retrieval strategies.\n\nIn addition to these method, I guess some ingenuity are needed to get gold medal.\n\nNumber of candidate are 100-1000 in most case.\n\n\n## Summary\nI summarize from the 1st to 10-th solutions which are informative and exciting.\nI really appreciate to the author of these articles.", "votes": null }, { "id": "1785944", "postDate": "05/12/2022 14:05:09", "content": "

Thank you for summarizing the top-10 place solutions in one helpful table! This is a great resource for anyone looking to learn from the best in this competition.

", "rawMarkdown": "Thank you for summarizing the top-10 place solutions in one helpful table! This is a great resource for anyone looking to learn from the best in this competition.", "votes": null }, { "id": "1786096", "postDate": "05/12/2022 15:49:19", "content": "

Thank you for your comment.
\nI’m happy to hear that.

", "rawMarkdown": "Thank you for your comment.\nI’m happy to hear that.", "votes": null }, { "id": "1790880", "postDate": "05/15/2022 12:47:34", "content": "

Thanks a lot. Saves my time, very helpful.

", "rawMarkdown": "Thanks a lot. Saves my time, very helpful.", "votes": null }, { "id": "1790965", "postDate": "05/15/2022 14:08:25", "content": "

Very helpful, thank you!

", "rawMarkdown": "Very helpful, thank you!", "votes": null }, { "id": "1796039", "postDate": "05/20/2022 12:11:30", "content": "

You're welcome.
\nIt's my pleasure.

", "rawMarkdown": "You're welcome.\nIt's my pleasure.", "votes": null }, { "id": "1796040", "postDate": "05/20/2022 12:12:15", "content": "

You're welcome.
\nI'm glad to hear that.

", "rawMarkdown": "You're welcome.\nI'm glad to hear that.", "votes": null }, { "id": "2342591", "postDate": "07/13/2023 02:14:33", "content": "

Thanks, it's very helpful. But the picture doesn't look clear.

", "rawMarkdown": "Thanks, it's very helpful. But the picture doesn't look clear.", "votes": null }, { "id": "2342593", "postDate": "07/13/2023 02:19:08", "content": "

Oh I changed my browser and the pircute shows clear. Thanks.

", "rawMarkdown": "Oh I changed my browser and the pircute shows clear. Thanks.", "votes": null } ], "comments": [ { "id": 1785944, "author_name": "", "author_url": "", "post_date": "05/12/2022 14:05:09", "content": "

Thank you for summarizing the top-10 place solutions in one helpful table! This is a great resource for anyone looking to learn from the best in this competition.

", "votes": null, "replies": [ { "id": 1786096, "author_name": "tetsuro731", "author_url": "", "post_date": "05/12/2022 15:49:19", "content": "

Thank you for your comment.
\nI’m happy to hear that.

", "votes": null, "replies": [] } ] }, { "id": 1790880, "author_name": "homoalways", "author_url": "", "post_date": "05/15/2022 12:47:34", "content": "

Thanks a lot. Saves my time, very helpful.

", "votes": null, "replies": [ { "id": 1796039, "author_name": "tetsuro731", "author_url": "", "post_date": "05/20/2022 12:11:30", "content": "

You're welcome.
\nIt's my pleasure.

", "votes": null, "replies": [] } ] }, { "id": 1790965, "author_name": "jacob34", "author_url": "", "post_date": "05/15/2022 14:08:25", "content": "

Very helpful, thank you!

", "votes": null, "replies": [ { "id": 1796040, "author_name": "tetsuro731", "author_url": "", "post_date": "05/20/2022 12:12:15", "content": "

You're welcome.
\nI'm glad to hear that.

", "votes": null, "replies": [] } ] }, { "id": 2342591, "author_name": "ghost0913", "author_url": "", "post_date": "07/13/2023 02:14:33", "content": "

Thanks, it's very helpful. But the picture doesn't look clear.

", "votes": null, "replies": [ { "id": 2342593, "author_name": "ghost0913", "author_url": "", "post_date": "07/13/2023 02:19:08", "content": "

Oh I changed my browser and the pircute shows clear. Thanks.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1785286": "Thanks to the organizers for this interesting competition.\nI read the following articles which are from 1st to 10-th place solutions.\nThese are very interesting so I summarize the following method to one table.\n\n## Links\n\n1. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324070\n2. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324197\n3. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324129\n4. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324094\n5. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324098\n6. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324075\n7. -\n8. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324185\n9. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324127\n10. https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324223\n\n## Table\n\n[![top10-table.png](https://i.postimg.cc/qMLsH2kH/top10-table.png)](https://postimg.cc/RNW6c66g)\n\n\nI can't find the articles from 7-th place so I don't fill the 7-th place in the table.\n\n## Ranking\nRanking strategies are similar among the top-10 teams.\nSurprisingly, it seem all teams use LightGBM for the ranking. \nAt 4-th place, DCN was used in addition to GBDT but it is said that \"The score of LightGBM is much better\".\nCatBoost was also used by several teams.\n\nFor training period, it seems that it's popular to use 4 - 9 weeks and last week was used for the validation.\nNumber of features are only written in some articles, but about 300 features are added to the models.\n\nRegarding the ensembles, GBDT was commonly used and MLP or different linear model are not used frequently.\n\n\n## Retrieval/Recall\n\nThere are a variety of Retrieval/Recall strategies comparing to the ranking.\nAccording to the 1st place article, it is said that \"the candidate generation strategy is the key to beyond limitation of accuracy, good feature engineering or modeling could be close to the limitation\".\n\nI think this sentence represents this competition and it's very interesting.\nAccording to the 5-th place article, its is said \"we spend most of the time to develop recall strategies\".\n\nAs already mentioned, there are a variety of retrieval method.\nTherefore, it's difficult to summarize in one table.\nFor more details, please see the individual great article.\n\nRegarding the retrieval method, the following method are commonly used.\n- repurchase\n- pupular item\n- item collaborative filterling\n\nThese methods had been mentioned in Discussion/Note during the competition.\nTherefore, I guess a lot of teams used these solutions to their retrieval strategies.\n\nIn addition to these method, I guess some ingenuity are needed to get gold medal.\n\nNumber of candidate are 100-1000 in most case.\n\n\n## Summary\nI summarize from the 1st to 10-th solutions which are informative and exciting.\nI really appreciate to the author of these articles.", "1785944": "Thank you for summarizing the top-10 place solutions in one helpful table! This is a great resource for anyone looking to learn from the best in this competition.", "1786096": "Thank you for your comment.\nI’m happy to hear that.", "1790880": "Thanks a lot. Saves my time, very helpful.", "1790965": "Very helpful, thank you!", "1796039": "You're welcome.\nIt's my pleasure.", "1796040": "You're welcome.\nI'm glad to hear that.", "2342591": "Thanks, it's very helpful. But the picture doesn't look clear.", "2342593": "Oh I changed my browser and the pircute shows clear. Thanks." }, "source": "meta" }