Thank you, Lucas and Carl, for organizing the competition. My solution consists of an ensemble approach using H2O and FLAML AutoML:
\nH2O AutoML:
\nFLAML AutoML:
\nFor feature engineering, I added one-hot encoding for each categorical feature, which was used alongside the original categorical feature across relevant datasets. Each model was trained for 120 minutes on a machine with 48 CPU cores and 380 GB of RAM over each dataset.
", "messages": [ { "id": "2794855", "postDate": "05/05/2024 14:42:42", "content": "Thank you, Lucas and Carl, for organizing the competition. My solution consists of an ensemble approach using H2O and FLAML AutoML:
\nH2O AutoML:
\nFLAML AutoML:
\nFor feature engineering, I added one-hot encoding for each categorical feature, which was used alongside the original categorical feature across relevant datasets. Each model was trained for 120 minutes on a machine with 48 CPU cores and 380 GB of RAM over each dataset.
", "rawMarkdown": "Thank you, Lucas and Carl, for organizing the competition. My solution consists of an ensemble approach using H2O and FLAML AutoML:\n\nH2O AutoML:\n- Utilized all available models, including stacking.\n- 5-fold cross-validation.\n\nFLAML AutoML:\n- Selected models: ['lgbm', 'xgboost', 'catboost', 'rf', 'extra_tree', 'xgb_limitdepth'] without ensembling or stacking.\n- 5-fold cross-validation.\n\nFor feature engineering, I added one-hot encoding for each categorical feature, which was used alongside the original categorical feature across relevant datasets. Each model was trained for 120 minutes on a machine with 48 CPU cores and 380 GB of RAM over each dataset.", "votes": null }, { "id": "2795004", "postDate": "05/05/2024 15:55:32", "content": "What was the total training time for all models? Did it actually use the full 2 hours per dataset?
", "rawMarkdown": "What was the total training time for all models? Did it actually use the full 2 hours per dataset?", "votes": null }, { "id": "2795056", "postDate": "05/05/2024 16:11:32", "content": "Thanks for sharing, hope to see you next time.
", "rawMarkdown": "Thanks for sharing, hope to see you next time.", "votes": null }, { "id": "2795167", "postDate": "05/05/2024 17:20:55", "content": "~236 hrs, yes each automl used the full 2 hrs per dataset.
", "rawMarkdown": "~236 hrs, yes each automl used the full 2 hrs per dataset.", "votes": null }, { "id": "2795201", "postDate": "05/05/2024 17:40:05", "content": "What is the total cost of such a job if I may ask? (Just curious)
", "rawMarkdown": "What is the total cost of such a job if I may ask? (Just curious)", "votes": null }, { "id": "2795321", "postDate": "05/05/2024 18:22:56", "content": "~838 USD. That's about $3.554 hourly.
", "rawMarkdown": "~838 USD. That's about $3.554 hourly.", "votes": null }, { "id": "2795448", "postDate": "05/05/2024 18:59:06", "content": "Wait, did you spend 1k on a community comp. with no reward ?
", "rawMarkdown": "Wait, did you spend 1k on a community comp. with no reward ?", "votes": null }, { "id": "2795524", "postDate": "05/05/2024 20:04:30", "content": "That's my personal machine. If I were renting, it would be comparable to a Google n2-standard-96 (96 vCPU, 48 core, 384 GB memory), and hey, it’s not too late for a reward! haha
", "rawMarkdown": "That's my personal machine. If I were renting, it would be comparable to a Google n2-standard-96 (96 vCPU, 48 core, 384 GB memory), and hey, it’s not too late for a reward! haha", "votes": null }, { "id": "2796573", "postDate": "05/06/2024 09:55:07", "content": "Ok I am curious about your machine. What is it ? Did you buy it for yourself?
\nAnd also:
\nI thought that I have a good machine with 32 threads, 64GB RAM and an RTX 3090… but that one blows me into the dust… 😅
", "rawMarkdown": "I thought that I have a good machine with 32 threads, 64GB RAM and an RTX 3090... but that one blows me into the dust... 😅", "votes": null }, { "id": "2797788", "postDate": "05/06/2024 22:56:38", "content": "It's a custom build with AMD EPYC 7552.
\nI joined it a bit late, missing the first day, but I am participating now. I am currently testing which of the three AutoML frameworks—H2O, FLAML, and Gluon—performs best, particularly in terms of robustness around stacking/ensembling versus finding the single best model. If my results are consistent and confirmed in next month's competition, it might be worth writing a note (not sure how much generalization we can have from synthetic data, given that the true source of the data generator is known), thanks for the suggestion anyway.
Thanks for sharing bro I appreciate it!! @arsanikzad
", "rawMarkdown": "Thanks for sharing bro I appreciate it!! @arsanikzad", "votes": null }, { "id": "3381531", "postDate": "12/24/2025 21:14:47", "content": "\nWhat are the main conclusions of this competition please? Was there a write-up?
\nCame across another paper \nhttps://arxiv.org/pdf/2106.03253
\nBut no discussion of the role of the features
\nThank you in advance!
", "rawMarkdown": "lucasmorin @carlmcbrideellis \n\nWhat are the main conclusions of this competition please? Was there a write-up?\n\nCame across another paper \nhttps://arxiv.org/pdf/2106.03253\n\nBut no discussion of the role of the features\n\nThank you in advance!", "votes": null } ], "comments": [ { "id": 2795004, "author_name": "thomasmeiner", "author_url": "", "post_date": "05/05/2024 15:55:32", "content": "What was the total training time for all models? Did it actually use the full 2 hours per dataset?
", "votes": null, "replies": [ { "id": 2795167, "author_name": "arsanikzad", "author_url": "", "post_date": "05/05/2024 17:20:55", "content": "~236 hrs, yes each automl used the full 2 hrs per dataset.
", "votes": null, "replies": [ { "id": 2795201, "author_name": "thomasmeiner", "author_url": "", "post_date": "05/05/2024 17:40:05", "content": "What is the total cost of such a job if I may ask? (Just curious)
", "votes": null, "replies": [ { "id": 2795321, "author_name": "arsanikzad", "author_url": "", "post_date": "05/05/2024 18:22:56", "content": "~838 USD. That's about $3.554 hourly.
", "votes": null, "replies": [ { "id": 2795448, "author_name": "lucasmorin", "author_url": "", "post_date": "05/05/2024 18:59:06", "content": "Wait, did you spend 1k on a community comp. with no reward ?
", "votes": null, "replies": [ { "id": 2795524, "author_name": "arsanikzad", "author_url": "", "post_date": "05/05/2024 20:04:30", "content": "That's my personal machine. If I were renting, it would be comparable to a Google n2-standard-96 (96 vCPU, 48 core, 384 GB memory), and hey, it’s not too late for a reward! haha
", "votes": null, "replies": [ { "id": 2796573, "author_name": "lucasmorin", "author_url": "", "post_date": "05/06/2024 09:55:07", "content": "Ok I am curious about your machine. What is it ? Did you buy it for yourself?
\nAnd also:
\nI thought that I have a good machine with 32 threads, 64GB RAM and an RTX 3090… but that one blows me into the dust… 😅
", "votes": null, "replies": [] }, { "id": 2797788, "author_name": "arsanikzad", "author_url": "", "post_date": "05/06/2024 22:56:38", "content": "It's a custom build with AMD EPYC 7552.
\nI joined it a bit late, missing the first day, but I am participating now. I am currently testing which of the three AutoML frameworks—H2O, FLAML, and Gluon—performs best, particularly in terms of robustness around stacking/ensembling versus finding the single best model. If my results are consistent and confirmed in next month's competition, it might be worth writing a note (not sure how much generalization we can have from synthetic data, given that the true source of the data generator is known), thanks for the suggestion anyway.
Thanks for sharing, hope to see you next time.
", "votes": null, "replies": [] }, { "id": 2800721, "author_name": "akifahiner", "author_url": "", "post_date": "05/08/2024 10:22:18", "content": "Thanks for sharing bro I appreciate it!! @arsanikzad
", "votes": null, "replies": [] }, { "id": 3381531, "author_name": "mihailturlakov", "author_url": "", "post_date": "12/24/2025 21:14:47", "content": "\nWhat are the main conclusions of this competition please? Was there a write-up?
\nCame across another paper \nhttps://arxiv.org/pdf/2106.03253
\nBut no discussion of the role of the features
\nThank you in advance!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2794855": "Thank you, Lucas and Carl, for organizing the competition. My solution consists of an ensemble approach using H2O and FLAML AutoML:\n\nH2O AutoML:\n- Utilized all available models, including stacking.\n- 5-fold cross-validation.\n\nFLAML AutoML:\n- Selected models: ['lgbm', 'xgboost', 'catboost', 'rf', 'extra_tree', 'xgb_limitdepth'] without ensembling or stacking.\n- 5-fold cross-validation.\n\nFor feature engineering, I added one-hot encoding for each categorical feature, which was used alongside the original categorical feature across relevant datasets. Each model was trained for 120 minutes on a machine with 48 CPU cores and 380 GB of RAM over each dataset.", "2795004": "What was the total training time for all models? Did it actually use the full 2 hours per dataset?", "2795056": "Thanks for sharing, hope to see you next time.", "2795167": "~236 hrs, yes each automl used the full 2 hrs per dataset.", "2795201": "What is the total cost of such a job if I may ask? (Just curious)", "2795321": "~838 USD. That's about $3.554 hourly.", "2795448": "Wait, did you spend 1k on a community comp. with no reward ?", "2795524": "That's my personal machine. If I were renting, it would be comparable to a Google n2-standard-96 (96 vCPU, 48 core, 384 GB memory), and hey, it’s not too late for a reward! haha", "2796573": "Ok I am curious about your machine. What is it ? Did you buy it for yourself?\n\nAnd also:\n- There is an AutoML challenge currently ongoing, what do you think about it ? did you participate ? \n- It seems you have an interesting result. Would you consider writing a note ? (and maybe submit it to the conference that organise the aforementioned AutoML challenge)", "2796636": "I thought that I have a good machine with 32 threads, 64GB RAM and an RTX 3090... but that one blows me into the dust... 😅", "2797788": "It's a custom build with AMD EPYC 7552.\nI joined it a bit late, missing the first day, but I am participating now. I am currently testing which of the three AutoML frameworks—H2O, FLAML, and Gluon—performs best, particularly in terms of robustness around stacking/ensembling versus finding the single best model. If my results are consistent and confirmed in next month's competition, it might be worth writing a note (not sure how much generalization we can have from synthetic data, given that the true source of the data generator is known), thanks for the suggestion anyway.", "2800721": "Thanks for sharing bro I appreciate it!! @arsanikzad", "3381531": "lucasmorin @carlmcbrideellis \n\nWhat are the main conclusions of this competition please? Was there a write-up?\n\nCame across another paper \nhttps://arxiv.org/pdf/2106.03253\n\nBut no discussion of the role of the features\n\nThank you in advance!" }, "source": "meta" }