On the final week we still have things to learn, there are amazing notebooks that shows how to make things easier.
\nThe notebooks I will like to explore are:
\nOn the final week we still have things to learn, there are amazing notebooks that shows how to make things easier.
\nThe notebooks I will like to explore are:
\n@pastorsoto thanks for mentioning me π
\nI will add my 5 cents. First of all, to figure out what are the improvements in my kernel, take a look on this posts (all of these tricks are used):
\nSecondly, if you are thinking about the AutoML solution - I'm the head of the LightAutoML project, the opensource AutoML library which can help you to create good models in the fast automatic way. @mukaseevru has created the example kernel how to use it on this competition. At the early stages of the competition I have used it on my PC and if you give it a bigger timeout and change f1-score metric to the logloss in Task object you can receive 0.19714 LB (leaderboard) score on full data and 0.19788 LB score on the 20% sample. After that I have created the augmented dataset shuffling players and teams to receive better score - I have created the 163 million rows train dataset with 92 features and 25 million rows test dataset. LightAutoML model was built on the such a huge amount of data in ~6*2 hours (using 32 CPU cores) for the both targets and the both predictions take ~18*2 minutes. This makes me to receive 0.19604 on LB, so LightAutoML can handle this amount of data without any problem.
Finally, I'm also going to create the topic with some important notes about the competition metric, the LogLoss - so stay tuned and follow me not to miss it π
\nHope this helps and good luck in your work!
\nAlex
", "rawMarkdown": "pastorsoto thanks for mentioning me π\n\nI will add my 5 cents. First of all, to figure out what are the improvements in my kernel, take a look on this posts (all of these tricks are used):\n- [Why Mish activation works here?](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/360084)\n- [Test time data augmentation (TTA) in the nutshell](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/359502)\n- [What validation should be used to decrease the metric gap and prevent overfitting?](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/359714)\n\nSecondly, if you are thinking about the AutoML solution - I'm the head of the [LightAutoML](https://github.com/sb-ai-lab/LightAutoML) project, the opensource AutoML library which can help you to create good models in the fast automatic way. @mukaseevru has created the [example kernel](https://www.kaggle.com/code/mukaseevru/tps-oct-22-lama-lightautoml-fe-sampling) how to use it on this competition. At the early stages of the competition I have used it on my PC and if you give it a bigger timeout and change f1-score metric to the logloss in `Task` object you can receive **0.19714** LB (leaderboard) score on full data and **0.19788** LB score on the 20% sample. After that I have created the augmented dataset shuffling players and teams to receive better score - I have created the 163 million rows train dataset with 92 features and 25 million rows test dataset. LightAutoML model was built on the such a huge amount of data in ~6\\*2 hours (using 32 CPU cores) for the both targets and the both predictions take ~18\\*2 minutes. This makes me to receive **0.19604** on LB, so LightAutoML can handle this amount of data without any problem.\n\nFinally, I'm also going to create the topic with some important notes about the competition metric, the LogLoss - so stay tuned and follow me not to miss it π\n\nHope this helps and good luck in your work!\n\nAlex", "votes": null }, { "id": "2001270", "postDate": "10/23/2022 21:36:34", "content": "Hi. That's amazing insight!!! Thank you so much. I will test your AutoML project, I think can give me a great benchmark, I will study your answer, it's really helpful for my learning journey.
", "rawMarkdown": "Hi. That's amazing insight!!! Thank you so much. I will test your AutoML project, I think can give me a great benchmark, I will study your answer, it's really helpful for my learning journey.", "votes": null } ], "comments": [ { "id": 2001161, "author_name": "alexryzhkov", "author_url": "", "post_date": "10/23/2022 19:58:42", "content": "@pastorsoto thanks for mentioning me π
\nI will add my 5 cents. First of all, to figure out what are the improvements in my kernel, take a look on this posts (all of these tricks are used):
\nSecondly, if you are thinking about the AutoML solution - I'm the head of the LightAutoML project, the opensource AutoML library which can help you to create good models in the fast automatic way. @mukaseevru has created the example kernel how to use it on this competition. At the early stages of the competition I have used it on my PC and if you give it a bigger timeout and change f1-score metric to the logloss in Task object you can receive 0.19714 LB (leaderboard) score on full data and 0.19788 LB score on the 20% sample. After that I have created the augmented dataset shuffling players and teams to receive better score - I have created the 163 million rows train dataset with 92 features and 25 million rows test dataset. LightAutoML model was built on the such a huge amount of data in ~6*2 hours (using 32 CPU cores) for the both targets and the both predictions take ~18*2 minutes. This makes me to receive 0.19604 on LB, so LightAutoML can handle this amount of data without any problem.
Finally, I'm also going to create the topic with some important notes about the competition metric, the LogLoss - so stay tuned and follow me not to miss it π
\nHope this helps and good luck in your work!
\nAlex
", "votes": null, "replies": [ { "id": 2001270, "author_name": "pastorsoto", "author_url": "", "post_date": "10/23/2022 21:36:34", "content": "Hi. That's amazing insight!!! Thank you so much. I will test your AutoML project, I think can give me a great benchmark, I will study your answer, it's really helpful for my learning journey.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2000814": "On the final week we still have things to learn, there are amazing notebooks that shows how to make things easier.\n\n## Missing values\n\n- XGBoost can handle missing values without imputation\n- Some imputation methods that I tried makes no big difference (I tried using a place on the field that is unlikely to add up to the probabilities of the team and binary feature to say if it is missing or not)\n\n## Save memory\n\n- Dask allows to create dataframe that speed up the process compare to pandas, which is a really great tool for our datasets.\n- use gc.collect() to allow clearing up the memory on each step, so it's something that we can take advantage between models. \n\n## Models\n\n- There are some AutoML that might work but not sure if they are able to handle big datasets\n- From this [notebook](https://www.kaggle.com/code/donatoriccio/how-to-load-21m-rows-in-1-minute-using-2-lines/notebook) by @donatoriccio I was able to get a better score without using any feature engineering or data manipulation\n\n## Next week plan\n\n- I have two notebooks, one for my own experiment and another one to test things from other notebooks\n- Try things learned on my own notebook \n- Experiment with things of public notebooks\n- My rule will be no copy notebooks, I'll be coding along and see if I am able to understand, if not I will try to search until I learn something, my final submission will be on my own notebook from the things learned.\n\nThe notebooks I will like to explore are:\n- [His visualizations are amazing](https://www.kaggle.com/code/sergiosaharovskiy/tps-oct-2022-viz-players-positions-animated/notebook) by @sergiosaharovskiy \n- [FastAI model](https://www.kaggle.com/code/paddykb/tps-2022-10-fastai) by @paddykb\n- [Amazing notebook with a lot of things to learn](https://www.kaggle.com/code/alexryzhkov/tps-2022-10-fastai-with-multistart-and-tta/notebook) by @alexryzhkov (This is an improvement of the @paddykb FastAI model, I want to explore how improvements are made and try to understand their process).", "2001161": "pastorsoto thanks for mentioning me π\n\nI will add my 5 cents. First of all, to figure out what are the improvements in my kernel, take a look on this posts (all of these tricks are used):\n- [Why Mish activation works here?](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/360084)\n- [Test time data augmentation (TTA) in the nutshell](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/359502)\n- [What validation should be used to decrease the metric gap and prevent overfitting?](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/359714)\n\nSecondly, if you are thinking about the AutoML solution - I'm the head of the [LightAutoML](https://github.com/sb-ai-lab/LightAutoML) project, the opensource AutoML library which can help you to create good models in the fast automatic way. @mukaseevru has created the [example kernel](https://www.kaggle.com/code/mukaseevru/tps-oct-22-lama-lightautoml-fe-sampling) how to use it on this competition. At the early stages of the competition I have used it on my PC and if you give it a bigger timeout and change f1-score metric to the logloss in `Task` object you can receive **0.19714** LB (leaderboard) score on full data and **0.19788** LB score on the 20% sample. After that I have created the augmented dataset shuffling players and teams to receive better score - I have created the 163 million rows train dataset with 92 features and 25 million rows test dataset. LightAutoML model was built on the such a huge amount of data in ~6\\*2 hours (using 32 CPU cores) for the both targets and the both predictions take ~18\\*2 minutes. This makes me to receive **0.19604** on LB, so LightAutoML can handle this amount of data without any problem.\n\nFinally, I'm also going to create the topic with some important notes about the competition metric, the LogLoss - so stay tuned and follow me not to miss it π\n\nHope this helps and good luck in your work!\n\nAlex", "2001270": "Hi. That's amazing insight!!! Thank you so much. I will test your AutoML project, I think can give me a great benchmark, I will study your answer, it's really helpful for my learning journey." }, "source": "meta" }