In this competition many decisions must be made based on the type of atoms. Modeling for each type proved to be very effective. I jumped from -0.85 to -0.924 with this strategy. (Kernel link). Now, I jumped to -1.0, using the same idea, but this time, adjusting the model parameters for each type of atom. I made a kernel that exemplifies how this technique can be used. The Kernel uses GridSearchCV to fetch the best parameters.
\n\nhttps://www.kaggle.com/adrianoavelar/gridsearch-for-eachtype-lb-1-0
", "messages": [ { "id": "568947", "postDate": "07/05/2019 18:25:27", "content": "In this competition many decisions must be made based on the type of atoms. Modeling for each type proved to be very effective. I jumped from -0.85 to -0.924 with this strategy. (Kernel link). Now, I jumped to -1.0, using the same idea, but this time, adjusting the model parameters for each type of atom. I made a kernel that exemplifies how this technique can be used. The Kernel uses GridSearchCV to fetch the best parameters.
\n\nhttps://www.kaggle.com/adrianoavelar/gridsearch-for-eachtype-lb-1-0
", "rawMarkdown": "In this competition many decisions must be made based on the type of atoms. Modeling for each type proved to be very effective. I jumped from -0.85 to -0.924 with this strategy. ([Kernel link](https://www.kaggle.com/adrianoavelar/eachtype)). Now, I jumped to -1.0, using the same idea, but this time, adjusting the model parameters for each type of atom. I made a kernel that exemplifies how this technique can be used. The Kernel uses GridSearchCV to fetch the best parameters.\n\nhttps://www.kaggle.com/adrianoavelar/gridsearch-for-eachtype-lb-1-0", "votes": null }, { "id": "568978", "postDate": "07/05/2019 19:39:51", "content": "Awesome. I am planning to do the same once I find a good set of features. \nHave you tried Bayesian methods for hp optimization (hyperopt for example)?
", "rawMarkdown": "Awesome. I am planning to do the same once I find a good set of features. \nHave you tried Bayesian methods for hp optimization ([**hyperopt**](https://www.kaggle.com/c/utexasdatasciencelab/discussion/24795) for example)?", "votes": null }, { "id": "571016", "postDate": "07/09/2019 04:03:38", "content": "Awesome. Interesting insights , i thought the max_depth would vary by type , but looks like it is constant at 15 ( overfitting ?? )
", "rawMarkdown": "Awesome. Interesting insights , i thought the max_depth would vary by type , but looks like it is constant at 15 ( overfitting ?? )", "votes": null }, { "id": "573217", "postDate": "07/12/2019 01:00:56", "content": "No. I'll check this out. Thanks @yassinealouini
", "rawMarkdown": "No. I'll check this out. Thanks @yassinealouini", "votes": null }, { "id": "573218", "postDate": "07/12/2019 01:03:42", "content": "Yeah! Because it's just an example kernel, I used only dummy params. But in my current kernel, 1JHC got the best CV with max_depth=5, you have to make your search with your own features.
", "rawMarkdown": "Yeah! Because it's just an example kernel, I used only dummy params. But in my current kernel, 1JHC got the best CV with max_depth=5, you have to make your search with your own features.", "votes": null } ], "comments": [ { "id": 568978, "author_name": "yassinealouini", "author_url": "", "post_date": "07/05/2019 19:39:51", "content": "Awesome. I am planning to do the same once I find a good set of features. \nHave you tried Bayesian methods for hp optimization (hyperopt for example)?
", "votes": null, "replies": [ { "id": 573217, "author_name": "adrianoavelar", "author_url": "", "post_date": "07/12/2019 01:00:56", "content": "No. I'll check this out. Thanks @yassinealouini
", "votes": null, "replies": [] } ] }, { "id": 571016, "author_name": "venkat555", "author_url": "", "post_date": "07/09/2019 04:03:38", "content": "Awesome. Interesting insights , i thought the max_depth would vary by type , but looks like it is constant at 15 ( overfitting ?? )
", "votes": null, "replies": [ { "id": 573218, "author_name": "adrianoavelar", "author_url": "", "post_date": "07/12/2019 01:03:42", "content": "Yeah! Because it's just an example kernel, I used only dummy params. But in my current kernel, 1JHC got the best CV with max_depth=5, you have to make your search with your own features.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "568947": "In this competition many decisions must be made based on the type of atoms. Modeling for each type proved to be very effective. I jumped from -0.85 to -0.924 with this strategy. ([Kernel link](https://www.kaggle.com/adrianoavelar/eachtype)). Now, I jumped to -1.0, using the same idea, but this time, adjusting the model parameters for each type of atom. I made a kernel that exemplifies how this technique can be used. The Kernel uses GridSearchCV to fetch the best parameters.\n\nhttps://www.kaggle.com/adrianoavelar/gridsearch-for-eachtype-lb-1-0", "568978": "Awesome. I am planning to do the same once I find a good set of features. \nHave you tried Bayesian methods for hp optimization ([**hyperopt**](https://www.kaggle.com/c/utexasdatasciencelab/discussion/24795) for example)?", "571016": "Awesome. Interesting insights , i thought the max_depth would vary by type , but looks like it is constant at 15 ( overfitting ?? )", "573217": "No. I'll check this out. Thanks @yassinealouini", "573218": "Yeah! Because it's just an example kernel, I used only dummy params. But in my current kernel, 1JHC got the best CV with max_depth=5, you have to make your search with your own features." }, "source": "meta" }