Do you know of any good resources on how ranking models work?
\nFor instance, how does breaking the data into sessions help the model train? How is this information used?
\nWhy the model seems to improve if I feed it more sessions EVEN if some of the sessions don't have any ground truth associated with them? (they are actual sessions but where nothing was clicked, for instance). I understand that negative examples can also be helpful, but how does the model make use of this information?
\nI am looking for resources that would build intuition on how these models work or alternatively that speak about how to improve the performance of these models.
\nThank you very much for your help 🙏
\nDo you know of any good resources on how ranking models work?
\nFor instance, how does breaking the data into sessions help the model train? How is this information used?
\nWhy the model seems to improve if I feed it more sessions EVEN if some of the sessions don't have any ground truth associated with them? (they are actual sessions but where nothing was clicked, for instance). I understand that negative examples can also be helpful, but how does the model make use of this information?
\nI am looking for resources that would build intuition on how these models work or alternatively that speak about how to improve the performance of these models.
\nThank you very much for your help 🙏
\nI 'm waiting for the experienced kagglers to answer your questions on this topic, so that I'll learn too.
", "rawMarkdown": "I 'm waiting for the experienced kagglers to answer your questions on this topic, so that I'll learn too.", "votes": null }, { "id": "2035208", "postDate": "11/18/2022 17:42:34", "content": "It is my understanding that there are 3 ways that GBT can rank
\nWhen choosing parameters for GBT, we can choose which method objective we want to use. Here is Wikipedia page about ranking here
Thank you very much, @cdeotte! 🙂 Very valuable information, as always 🙂
\nNow the question is how do various frameworks expose these options and whether they do 🤔 I only played around very briefly with lgbm in this competition so far, I believe it only supports lambdarank as the objective.
Not too much information on that and the documentation of the library is not great, but I found this blog post quite informative.
\nNext steps would probably be xgboost and possibly even catboost! 🙂
You can go for voting or bagging
\nalso do check my profile
try this paper @radek1 ,https://www.microsoft.com/en-us/research/uploads/prod/2016/02/MSR-TR-2010-82.pdf this gives clear info
", "rawMarkdown": "try this paper @radek1 ,https://www.microsoft.com/en-us/research/uploads/prod/2016/02/MSR-TR-2010-82.pdf this gives clear info", "votes": null }, { "id": "2041945", "postDate": "11/24/2022 10:16:21", "content": "Thank you for the recommendation, @satyanarayanam! 🙌🙂
", "rawMarkdown": "Thank you for the recommendation, @satyanarayanam! 🙌🙂", "votes": null } ], "comments": [ { "id": 2034655, "author_name": "mpwolke", "author_url": "", "post_date": "11/18/2022 10:22:56", "content": "I 'm waiting for the experienced kagglers to answer your questions on this topic, so that I'll learn too.
", "votes": null, "replies": [] }, { "id": 2035208, "author_name": "cdeotte", "author_url": "", "post_date": "11/18/2022 17:42:34", "content": "It is my understanding that there are 3 ways that GBT can rank
\nWhen choosing parameters for GBT, we can choose which method objective we want to use. Here is Wikipedia page about ranking here
Thank you very much, @cdeotte! 🙂 Very valuable information, as always 🙂
\nNow the question is how do various frameworks expose these options and whether they do 🤔 I only played around very briefly with lgbm in this competition so far, I believe it only supports lambdarank as the objective.
Not too much information on that and the documentation of the library is not great, but I found this blog post quite informative.
\nNext steps would probably be xgboost and possibly even catboost! 🙂
You can go for voting or bagging
\nalso do check my profile
try this paper @radek1 ,https://www.microsoft.com/en-us/research/uploads/prod/2016/02/MSR-TR-2010-82.pdf this gives clear info
", "votes": null, "replies": [ { "id": 2041945, "author_name": "radek1", "author_url": "", "post_date": "11/24/2022 10:16:21", "content": "Thank you for the recommendation, @satyanarayanam! 🙌🙂
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2031909": "Do you know of any good resources on how ranking models work?\n\nFor instance, how does breaking the data into sessions help the model train? How is this information used?\n\nWhy the model seems to improve if I feed it more sessions EVEN if some of the sessions don't have any ground truth associated with them? (they are actual sessions but where nothing was clicked, for instance). I understand that negative examples can also be helpful, but how does the model make use of this information?\n\nI am looking for resources that would build intuition on how these models work or alternatively that speak about how to improve the performance of these models.\n\nThank you very much for your help 🙏\n\n### Other resources you might find useful:\n\n* [💡 [2 methods] How-to ensemble predictions 🏅🏅🏅](https://www.kaggle.com/code/radek1/2-methods-how-to-ensemble-predictions)\n* [local validation tracks public LB perfecty -- here is the setup](https://www.kaggle.com/competitions/otto-recommender-system/discussion/364991)\n* [💡 For my friends from Twitter and LinkedIn -- here is how to dive into this competition 🐳](https://www.kaggle.com/competitions/otto-recommender-system/discussion/368560)\n* [Full dataset processed to CSV/parquet files with optimized memory footprint](https://www.kaggle.com/competitions/otto-recommender-system/discussion/363843)\n* [co-visitation matrix - simplified, imprvd logic 🔥](https://www.kaggle.com/code/radek1/co-visitation-matrix-simplified-imprvd-logic)\n* [💡 Word2Vec How-to [training and submission]🚀🚀🚀](https://www.kaggle.com/code/radek1/word2vec-how-to-training-and-submission)", "2034655": "I 'm waiting for the experienced kagglers to answer your questions on this topic, so that I'll learn too.", "2035208": "It is my understanding that there are 3 ways that GBT can rank\n* pointwise - learns regression on each item individually. Then sorts predicted scores\n* pairwise - learns binary classification on each pair of items. Then converts predictions into ranked list\n* listwise - uses complex algorithms that analyze entire list at once. There are many different algorithms and research papers.\n\nWhen choosing parameters for GBT, we can choose which method `objective` we want to use. Here is Wikipedia page about ranking [here][1]\n\n[1]: https://en.wikipedia.org/wiki/Learning_to_rank", "2035360": "Thank you very much, @cdeotte! 🙂 Very valuable information, as always 🙂\n\nNow the question is how do various frameworks expose these options and whether they do 🤔 I only played around very briefly with lgbm in this competition so far, I believe it only supports `lambdarank` as the objective.\n\nNot too much information on that and the documentation of the library is not great, but I found [this blog post quite informative](https://ffineis.github.io/blog/2021/05/01/lambdarank-lightgbm.html).\n\nNext steps would probably be `xgboost` and possibly even `catboost`! 🙂", "2035696": "You can go for voting or bagging\nalso do check my profile", "2041941": "try this paper @radek1 ,https://www.microsoft.com/en-us/research/uploads/prod/2016/02/MSR-TR-2010-82.pdf this gives clear info", "2041945": "Thank you for the recommendation, @satyanarayanam! 🙌🙂" }, "source": "meta" }