MinMax ensemble was one of the most popular notebooks at this competition. Ensembles of public decisions in this competition were normal and MinMax was one of the most popular ways to do it. However, I was skeptical. I've seen a lot of people take a medal just by averaging public submissions. Are there those who got a good score on private LB using exactly the MinMax ensemble?
", "messages": [ { "id": "975277", "postDate": "08/18/2020 08:19:29", "content": "MinMax ensemble was one of the most popular notebooks at this competition. Ensembles of public decisions in this competition were normal and MinMax was one of the most popular ways to do it. However, I was skeptical. I've seen a lot of people take a medal just by averaging public submissions. Are there those who got a good score on private LB using exactly the MinMax ensemble?
", "rawMarkdown": "MinMax ensemble was one of the most popular notebooks at this competition. Ensembles of public decisions in this competition were normal and MinMax was one of the most popular ways to do it. However, I was skeptical. I've seen a lot of people take a medal just by averaging public submissions. Are there those who got a good score on private LB using exactly the MinMax ensemble?", "votes": null }, { "id": "975304", "postDate": "08/18/2020 08:32:54", "content": "Definitely not
", "rawMarkdown": "Definitely not", "votes": null }, { "id": "975317", "postDate": "08/18/2020 08:39:42", "content": "Any of my 'fancy' ensembles scored worse than simple ensembles, although Chris has posted a search method for optimising weights.
\nI found three methods worked well:
\nNo.. Mean ensembling and rank ensembling were working much better in this competition (atleast in our submissions).
", "rawMarkdown": "No.. Mean ensembling and rank ensembling were working much better in this competition (atleast in our submissions).", "votes": null }, { "id": "978090", "postDate": "08/19/2020 23:41:59", "content": "It was worse than simple averaging in most cases. Weighted ensembling based on the LB AUC score, or sensitivity worked better for me.
", "rawMarkdown": "It was worse than simple averaging in most cases. Weighted ensembling based on the LB AUC score, or sensitivity worked better for me.", "votes": null }, { "id": "979123", "postDate": "08/20/2020 16:22:52", "content": "Applying MinMax ensemble on our set of models produces a worse result in both CV and LB.
\nAt the same time, our best solution combined top-k rank average and stacking, which performed better than simple blends on both public and private LB.
", "rawMarkdown": "Applying MinMax ensemble on our set of models produces a worse result in both CV and LB. \n\nAt the same time, our best solution combined top-k rank average and stacking, which performed better than simple blends on both public and private LB.", "votes": null } ], "comments": [ { "id": 975304, "author_name": "vicioussong", "author_url": "", "post_date": "08/18/2020 08:32:54", "content": "Definitely not
", "votes": null, "replies": [] }, { "id": 975317, "author_name": "fchmiel", "author_url": "", "post_date": "08/18/2020 08:39:42", "content": "Any of my 'fancy' ensembles scored worse than simple ensembles, although Chris has posted a search method for optimising weights.
\nI found three methods worked well:
\nNo.. Mean ensembling and rank ensembling were working much better in this competition (atleast in our submissions).
", "votes": null, "replies": [] }, { "id": 978090, "author_name": "antaresnyc", "author_url": "", "post_date": "08/19/2020 23:41:59", "content": "It was worse than simple averaging in most cases. Weighted ensembling based on the LB AUC score, or sensitivity worked better for me.
", "votes": null, "replies": [] }, { "id": 979123, "author_name": "kozodoi", "author_url": "", "post_date": "08/20/2020 16:22:52", "content": "Applying MinMax ensemble on our set of models produces a worse result in both CV and LB.
\nAt the same time, our best solution combined top-k rank average and stacking, which performed better than simple blends on both public and private LB.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "975277": "MinMax ensemble was one of the most popular notebooks at this competition. Ensembles of public decisions in this competition were normal and MinMax was one of the most popular ways to do it. However, I was skeptical. I've seen a lot of people take a medal just by averaging public submissions. Are there those who got a good score on private LB using exactly the MinMax ensemble?", "975304": "Definitely not", "975317": "Any of my 'fancy' ensembles scored worse than simple ensembles, although Chris has posted a search method for optimising weights.\n\nI found three methods worked well:\n\n- Straight up average\n- Weighted averaging with model weights equal to model OOF AUC\n- Weighted average with model weights equal to either 2 or 1.", "975452": "No.. Mean ensembling and rank ensembling were working much better in this competition (atleast in our submissions).", "978090": "It was worse than simple averaging in most cases. Weighted ensembling based on the LB AUC score, or sensitivity worked better for me.", "979123": "Applying MinMax ensemble on our set of models produces a worse result in both CV and LB. \n\nAt the same time, our best solution combined top-k rank average and stacking, which performed better than simple blends on both public and private LB." }, "source": "meta" }