{ "id": 367463, "title": "What happened so far? Full list of resources and discussions for a quick start to the competition πŸš€", "url": "/competitions/otto-recommender-system/discussion/367463", "author_name": "", "post_date": "2022-11-20T19:13:45.030778Z", "votes": 21, "comment_count": 5, "views": 0, "content": "

Hi everyone,

\n

It's been a while since the competition started so I would like to share the all resources that have been discussed/shared so far for people who join the party later:

\n

Discussions:

\n
    \n
  1. πŸ“ˆ What do we know so far? ⚑Summary with links to relevant resources by @radek1. Radek summarized the first weeks of the competition quite well already + with additional more recent resources.
    2. \n
  2. A top-down perspective on the current metric values by @narsil. A discussion about how sessions are truncated and how the dataset is created (also confirmed by the organizers).
    3. \n
  3. Yes, We Can Use Test Data Leakage by @cdeotte. A discussion where Chris shows us why it makes sense to use the test data in the training and how (in the second notebook below).
    4. \n
  4. Recommendation Systems for Large Datasets by @ravishah1. Ravi points out important parts of large scale recommender systems: candidate generation and ranking. He also gives additional directions on how to tackle those problems.
    5. \n
  5. πŸ’‘A robust local validation framework πŸš€πŸš€πŸš€ by @radek1. Radek shows us how to create a robust validation pipeline by using the last week of the train data as hold-out data without any inclusion of original test data.
    6. \n
  6. local validation tracks public LB perfecty -- here is the setup by @radek1. Radek shares his local validation setup with code examples as explained above.
    7. \n
  7. πŸ’‘ What is the co-visitation matrix, really? by @radek1. Radek explains what co-visitation matrix is and how it is useful for this competition so far.
    8. \n
  8. Locate Real Users and Real Sessions EDA by @cdeotte. With the accompanying notebook 4 - Chris shows us how to detect the real user sessions given the full month of data.
    9. \n
  9. [Starter pack] LGBMRanker with polars πŸš€πŸš€πŸš€ by @radek1.
    10. \n
  10. [Starter Pack] Matrix Factorization [Pytorch + Merlin Dataloader] πŸš€πŸš€πŸš€ by @radek1. In this discussion and accompanying notebook 6 - Radek shows us how to train a matrix factorization model using Pytorch - Merlin's Data Loader.
    11. \n
\n

Notebooks:

\n
    \n
  1. Co-visitation Matrix by vslaykovsky. Leveraging the idea of frequently viewed and bought items by creating a co-visitation matrix.
    2. \n
  2. Test Data Leak - LB Boost by @cdeotte. Chris demonstrates how to use the test data in the training. It also has a good track of experiment logs.
    3. \n
  3. Candidate ReRank Model - [LB 0.575] by @cdeotte. In this notebook, Chris shows us how to generate the candidates (most popular clicks, matrices, etc.) and rank them using hand-crafted features.
    4. \n
  4. Time Series EDA - Users and Real Sessions by @cdeotte.
    5. \n
  5. πŸ’‘ [polars] Proof of concept: LGBM RankerπŸ§ͺπŸ§ͺπŸ§ͺ by @radek1. In this notebook, Radek shows us how to build an LGBMRanker model.
    6. \n
  6. πŸ’‘Matrix Factorization [PyTorch+Merlin Dataloader]\n by @radek1.
    7. \n
  7. Pretraining with Merlin's Transformers4RecπŸ§™ by @theoviel. In this notebook, Theo shows us how to use Merlin's Transformers4Rec library to pretrain an XLNet language model using the next item prediction task.
    8. \n
  8. Compute Validation Score - [CV 565] by @cdeotte. In this notebook, Chris shows us how to compute the validation score given Radek's schema replicating the original test period.
    9. \n
", "messages": [ { "id": "2037622", "postDate": "11/20/2022 19:13:45", "content": "

Hi everyone,

\n

It's been a while since the competition started so I would like to share the all resources that have been discussed/shared so far for people who join the party later:

\n

Discussions:

\n
    \n
  1. πŸ“ˆ What do we know so far? ⚑Summary with links to relevant resources by @radek1. Radek summarized the first weeks of the competition quite well already + with additional more recent resources.
    2. \n
  2. A top-down perspective on the current metric values by @narsil. A discussion about how sessions are truncated and how the dataset is created (also confirmed by the organizers).
    3. \n
  3. Yes, We Can Use Test Data Leakage by @cdeotte. A discussion where Chris shows us why it makes sense to use the test data in the training and how (in the second notebook below).
    4. \n
  4. Recommendation Systems for Large Datasets by @ravishah1. Ravi points out important parts of large scale recommender systems: candidate generation and ranking. He also gives additional directions on how to tackle those problems.
    5. \n
  5. πŸ’‘A robust local validation framework πŸš€πŸš€πŸš€ by @radek1. Radek shows us how to create a robust validation pipeline by using the last week of the train data as hold-out data without any inclusion of original test data.
    6. \n
  6. local validation tracks public LB perfecty -- here is the setup by @radek1. Radek shares his local validation setup with code examples as explained above.
    7. \n
  7. πŸ’‘ What is the co-visitation matrix, really? by @radek1. Radek explains what co-visitation matrix is and how it is useful for this competition so far.
    8. \n
  8. Locate Real Users and Real Sessions EDA by @cdeotte. With the accompanying notebook 4 - Chris shows us how to detect the real user sessions given the full month of data.
    9. \n
  9. [Starter pack] LGBMRanker with polars πŸš€πŸš€πŸš€ by @radek1.
    10. \n
  10. [Starter Pack] Matrix Factorization [Pytorch + Merlin Dataloader] πŸš€πŸš€πŸš€ by @radek1. In this discussion and accompanying notebook 6 - Radek shows us how to train a matrix factorization model using Pytorch - Merlin's Data Loader.
    11. \n
\n

Notebooks:

\n
    \n
  1. Co-visitation Matrix by vslaykovsky. Leveraging the idea of frequently viewed and bought items by creating a co-visitation matrix.
    2. \n
  2. Test Data Leak - LB Boost by @cdeotte. Chris demonstrates how to use the test data in the training. It also has a good track of experiment logs.
    3. \n
  3. Candidate ReRank Model - [LB 0.575] by @cdeotte. In this notebook, Chris shows us how to generate the candidates (most popular clicks, matrices, etc.) and rank them using hand-crafted features.
    4. \n
  4. Time Series EDA - Users and Real Sessions by @cdeotte.
    5. \n
  5. πŸ’‘ [polars] Proof of concept: LGBM RankerπŸ§ͺπŸ§ͺπŸ§ͺ by @radek1. In this notebook, Radek shows us how to build an LGBMRanker model.
    6. \n
  6. πŸ’‘Matrix Factorization [PyTorch+Merlin Dataloader]\n by @radek1.
    7. \n
  7. Pretraining with Merlin's Transformers4RecπŸ§™ by @theoviel. In this notebook, Theo shows us how to use Merlin's Transformers4Rec library to pretrain an XLNet language model using the next item prediction task.
    8. \n
  8. Compute Validation Score - [CV 565] by @cdeotte. In this notebook, Chris shows us how to compute the validation score given Radek's schema replicating the original test period.
    9. \n
", "rawMarkdown": "Hi everyone,\n\nIt's been a while since the competition started so I would like to share the all resources that have been discussed/shared so far for people who join the party later:\n\n#### Discussions:\n\n1. [πŸ“ˆ What do we know so far? ⚑Summary with links to relevant resources](https://www.kaggle.com/competitions/otto-recommender-system/discussion/364062) by [@radek1](https://www.kaggle.com/radek1). Radek summarized the first weeks of the competition quite well already + with additional more recent resources. \n2. [A top-down perspective on the current metric values](https://www.kaggle.com/competitions/otto-recommender-system/discussion/363874) by [@narsil](https://www.kaggle.com/narsil). A discussion about how sessions are truncated and how the dataset is created (also confirmed by the organizers).\n3. [Yes, We Can Use Test Data Leakage](https://www.kaggle.com/competitions/otto-recommender-system/discussion/363939) by [@cdeotte](https://www.kaggle.com/cdeotte). A discussion where Chris shows us why it makes sense to use the test data in the training and how (in the second notebook below).\n4. [Recommendation Systems for Large Datasets](https://www.kaggle.com/competitions/otto-recommender-system/discussion/364721) by [@ravishah1](https://www.kaggle.com/ravishah1). Ravi points out important parts of large scale recommender systems: candidate generation and ranking. He also gives additional directions on how to tackle those problems. \n5. [πŸ’‘A robust local validation framework πŸš€πŸš€πŸš€](https://www.kaggle.com/competitions/otto-recommender-system/discussion/364216) by [@radek1](https://www.kaggle.com/radek1). Radek shows us how to create a robust validation pipeline by using the last week of the train data as hold-out data without any inclusion of original test data.\n6. [local validation tracks public LB perfecty -- here is the setup](https://www.kaggle.com/competitions/otto-recommender-system/discussion/364991) by [@radek1](https://www.kaggle.com/radek1). Radek shares his local validation setup with code examples as explained above.\n7. [πŸ’‘ What is the co-visitation matrix, really?](https://www.kaggle.com/competitions/otto-recommender-system/discussion/365358) by [@radek1](https://www.kaggle.com/radek1). Radek explains what co-visitation matrix is and how it is useful for this competition so far.\n8. [Locate Real Users and Real Sessions EDA](https://www.kaggle.com/competitions/otto-recommender-system/discussion/366138) by [@cdeotte](https://www.kaggle.com/cdeotte). With the accompanying notebook 4 - Chris shows us how to detect the real user sessions given the full month of data.\n9. [[Starter pack] LGBMRanker with polars πŸš€πŸš€πŸš€](https://www.kaggle.com/competitions/otto-recommender-system/discussion/366194) by [@radek1](https://www.kaggle.com/radek1).\n10. [[Starter Pack] Matrix Factorization [Pytorch + Merlin Dataloader] πŸš€πŸš€πŸš€](https://www.kaggle.com/competitions/otto-recommender-system/discussion/366893) by [@radek1](https://www.kaggle.com/radek1). In this discussion and accompanying notebook 6 - Radek shows us how to train a matrix factorization model using Pytorch - Merlin's Data Loader.\n\n#### Notebooks:\n\n1. [Co-visitation Matrix](https://www.kaggle.com/code/vslaykovsky/co-visitation-matrix) by [vslaykovsky](https://www.kaggle.com/vslaykovsky). Leveraging the idea of frequently viewed and bought items by creating a co-visitation matrix.\n2. [Test Data Leak - LB Boost](https://www.kaggle.com/code/cdeotte/test-data-leak-lb-boost) by [@cdeotte](https://www.kaggle.com/cdeotte). Chris demonstrates how to use the test data in the training. It also has a good track of experiment logs.\n3. [Candidate ReRank Model - [LB 0.575]](https://www.kaggle.com/code/cdeotte/candidate-rerank-model-lb-0-575?scriptVersionId=111214204) by [@cdeotte](https://www.kaggle.com/cdeotte). In this notebook, Chris shows us how to generate the candidates (most popular clicks, matrices, etc.) and rank them using hand-crafted features.\n4. [Time Series EDA - Users and Real Sessions](https://www.kaggle.com/code/cdeotte/time-series-eda-users-and-real-sessions) by [@cdeotte](https://www.kaggle.com/cdeotte).\n5. [πŸ’‘ [polars] Proof of concept: LGBM RankerπŸ§ͺπŸ§ͺπŸ§ͺ](https://www.kaggle.com/code/radek1/polars-proof-of-concept-lgbm-ranker/notebook) by [@radek1](https://www.kaggle.com/radek1). In this notebook, Radek shows us how to build an LGBMRanker model.\n6. [πŸ’‘Matrix Factorization [PyTorch+Merlin Dataloader]\n](https://www.kaggle.com/code/radek1/matrix-factorization-pytorch-merlin-dataloader) by [@radek1](https://www.kaggle.com/radek1).\n7. [Pretraining with Merlin's Transformers4RecπŸ§™](https://www.kaggle.com/code/theoviel/pretraining-with-merlin-s-transformers4rec) by [@theoviel](https://www.kaggle.com/theoviel). In this notebook, Theo shows us how to use Merlin's Transformers4Rec library to pretrain an XLNet language model using the next item prediction task.\n8. [Compute Validation Score - [CV 565]](https://www.kaggle.com/code/cdeotte/compute-validation-score-cv-565) by [@cdeotte](https://www.kaggle.com/cdeotte). In this notebook, Chris shows us how to compute the validation score given Radek's schema replicating the original test period.", "votes": null }, { "id": "2037651", "postDate": "11/20/2022 19:30:04", "content": "

Nice summary

", "rawMarkdown": "Nice summary", "votes": null }, { "id": "2037664", "postDate": "11/20/2022 19:49:49", "content": "

Thanks Chris! Thanks for the great resources as always. Half of it directly comes from your discussions/notebooks πŸ˜…

", "rawMarkdown": "Thanks Chris! Thanks for the great resources as always. Half of it directly comes from your discussions/notebooks πŸ˜…", "votes": null }, { "id": "2037682", "postDate": "11/20/2022 20:17:06", "content": "

Thank you, i'm questioning whose solution is better : Work2Vec or Analyse Matrix TF-IDF. If the order of chose of product in one session is important, I'm thinking that work2vec is better than analyse tf-idf.

", "rawMarkdown": "Thank you, i'm questioning whose solution is better : Work2Vec or Analyse Matrix TF-IDF. If the order of chose of product in one session is important, I'm thinking that work2vec is better than analyse tf-idf.", "votes": null }, { "id": "2037696", "postDate": "11/20/2022 20:33:34", "content": "

I would suggest using both and many additional others for candidate generation. Then use a ranking model for selection of final 20

", "rawMarkdown": "I would suggest using both and many additional others for candidate generation. Then use a ranking model for selection of final 20", "votes": null }, { "id": "2037720", "postDate": "11/20/2022 21:36:37", "content": "

Thanks for putting this up here. Nice to have a list.

", "rawMarkdown": "Thanks for putting this up here. Nice to have a list.", "votes": null } ], "comments": [ { "id": 2037651, "author_name": "cdeotte", "author_url": "", "post_date": "11/20/2022 19:30:04", "content": "

Nice summary

", "votes": null, "replies": [ { "id": 2037664, "author_name": "snnclsr", "author_url": "", "post_date": "11/20/2022 19:49:49", "content": "

Thanks Chris! Thanks for the great resources as always. Half of it directly comes from your discussions/notebooks πŸ˜…

", "votes": null, "replies": [] } ] }, { "id": 2037682, "author_name": "lgregory", "author_url": "", "post_date": "11/20/2022 20:17:06", "content": "

Thank you, i'm questioning whose solution is better : Work2Vec or Analyse Matrix TF-IDF. If the order of chose of product in one session is important, I'm thinking that work2vec is better than analyse tf-idf.

", "votes": null, "replies": [ { "id": 2037696, "author_name": "cdeotte", "author_url": "", "post_date": "11/20/2022 20:33:34", "content": "

I would suggest using both and many additional others for candidate generation. Then use a ranking model for selection of final 20

", "votes": null, "replies": [] } ] }, { "id": 2037720, "author_name": "megan3", "author_url": "", "post_date": "11/20/2022 21:36:37", "content": "

Thanks for putting this up here. Nice to have a list.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2037622": "Hi everyone,\n\nIt's been a while since the competition started so I would like to share the all resources that have been discussed/shared so far for people who join the party later:\n\n#### Discussions:\n\n1. [πŸ“ˆ What do we know so far? ⚑Summary with links to relevant resources](https://www.kaggle.com/competitions/otto-recommender-system/discussion/364062) by [@radek1](https://www.kaggle.com/radek1). Radek summarized the first weeks of the competition quite well already + with additional more recent resources. \n2. [A top-down perspective on the current metric values](https://www.kaggle.com/competitions/otto-recommender-system/discussion/363874) by [@narsil](https://www.kaggle.com/narsil). A discussion about how sessions are truncated and how the dataset is created (also confirmed by the organizers).\n3. [Yes, We Can Use Test Data Leakage](https://www.kaggle.com/competitions/otto-recommender-system/discussion/363939) by [@cdeotte](https://www.kaggle.com/cdeotte). A discussion where Chris shows us why it makes sense to use the test data in the training and how (in the second notebook below).\n4. [Recommendation Systems for Large Datasets](https://www.kaggle.com/competitions/otto-recommender-system/discussion/364721) by [@ravishah1](https://www.kaggle.com/ravishah1). Ravi points out important parts of large scale recommender systems: candidate generation and ranking. He also gives additional directions on how to tackle those problems. \n5. [πŸ’‘A robust local validation framework πŸš€πŸš€πŸš€](https://www.kaggle.com/competitions/otto-recommender-system/discussion/364216) by [@radek1](https://www.kaggle.com/radek1). Radek shows us how to create a robust validation pipeline by using the last week of the train data as hold-out data without any inclusion of original test data.\n6. [local validation tracks public LB perfecty -- here is the setup](https://www.kaggle.com/competitions/otto-recommender-system/discussion/364991) by [@radek1](https://www.kaggle.com/radek1). Radek shares his local validation setup with code examples as explained above.\n7. [πŸ’‘ What is the co-visitation matrix, really?](https://www.kaggle.com/competitions/otto-recommender-system/discussion/365358) by [@radek1](https://www.kaggle.com/radek1). Radek explains what co-visitation matrix is and how it is useful for this competition so far.\n8. [Locate Real Users and Real Sessions EDA](https://www.kaggle.com/competitions/otto-recommender-system/discussion/366138) by [@cdeotte](https://www.kaggle.com/cdeotte). With the accompanying notebook 4 - Chris shows us how to detect the real user sessions given the full month of data.\n9. [[Starter pack] LGBMRanker with polars πŸš€πŸš€πŸš€](https://www.kaggle.com/competitions/otto-recommender-system/discussion/366194) by [@radek1](https://www.kaggle.com/radek1).\n10. [[Starter Pack] Matrix Factorization [Pytorch + Merlin Dataloader] πŸš€πŸš€πŸš€](https://www.kaggle.com/competitions/otto-recommender-system/discussion/366893) by [@radek1](https://www.kaggle.com/radek1). In this discussion and accompanying notebook 6 - Radek shows us how to train a matrix factorization model using Pytorch - Merlin's Data Loader.\n\n#### Notebooks:\n\n1. [Co-visitation Matrix](https://www.kaggle.com/code/vslaykovsky/co-visitation-matrix) by [vslaykovsky](https://www.kaggle.com/vslaykovsky). Leveraging the idea of frequently viewed and bought items by creating a co-visitation matrix.\n2. [Test Data Leak - LB Boost](https://www.kaggle.com/code/cdeotte/test-data-leak-lb-boost) by [@cdeotte](https://www.kaggle.com/cdeotte). Chris demonstrates how to use the test data in the training. It also has a good track of experiment logs.\n3. [Candidate ReRank Model - [LB 0.575]](https://www.kaggle.com/code/cdeotte/candidate-rerank-model-lb-0-575?scriptVersionId=111214204) by [@cdeotte](https://www.kaggle.com/cdeotte). In this notebook, Chris shows us how to generate the candidates (most popular clicks, matrices, etc.) and rank them using hand-crafted features.\n4. [Time Series EDA - Users and Real Sessions](https://www.kaggle.com/code/cdeotte/time-series-eda-users-and-real-sessions) by [@cdeotte](https://www.kaggle.com/cdeotte).\n5. [πŸ’‘ [polars] Proof of concept: LGBM RankerπŸ§ͺπŸ§ͺπŸ§ͺ](https://www.kaggle.com/code/radek1/polars-proof-of-concept-lgbm-ranker/notebook) by [@radek1](https://www.kaggle.com/radek1). In this notebook, Radek shows us how to build an LGBMRanker model.\n6. [πŸ’‘Matrix Factorization [PyTorch+Merlin Dataloader]\n](https://www.kaggle.com/code/radek1/matrix-factorization-pytorch-merlin-dataloader) by [@radek1](https://www.kaggle.com/radek1).\n7. [Pretraining with Merlin's Transformers4RecπŸ§™](https://www.kaggle.com/code/theoviel/pretraining-with-merlin-s-transformers4rec) by [@theoviel](https://www.kaggle.com/theoviel). In this notebook, Theo shows us how to use Merlin's Transformers4Rec library to pretrain an XLNet language model using the next item prediction task.\n8. [Compute Validation Score - [CV 565]](https://www.kaggle.com/code/cdeotte/compute-validation-score-cv-565) by [@cdeotte](https://www.kaggle.com/cdeotte). In this notebook, Chris shows us how to compute the validation score given Radek's schema replicating the original test period.", "2037651": "Nice summary", "2037664": "Thanks Chris! Thanks for the great resources as always. Half of it directly comes from your discussions/notebooks πŸ˜…", "2037682": "Thank you, i'm questioning whose solution is better : Work2Vec or Analyse Matrix TF-IDF. If the order of chose of product in one session is important, I'm thinking that work2vec is better than analyse tf-idf.", "2037696": "I would suggest using both and many additional others for candidate generation. Then use a ranking model for selection of final 20", "2037720": "Thanks for putting this up here. Nice to have a list." }, "source": "meta" }