{ "id": 238722, "title": "Dark magic of aliens", "url": "/competitions/seti-breakthrough-listen/discussion/238722", "author_name": "ITK8191", "post_date": "2021-05-13T07:46:14.460000", "votes": 39, "comment_count": 2, "views": 0, "content": "

This competition has the AUC evaluation metric and ensemble notebook is very powerful.
\nThis reminds me of the RANZCR CLiP - Catheter and Line Position Challenge competition.
\nBy the way, @hengck23 posted a discussion titled one dark magic trick.
\nit means that in the ensemble notes, just before weighting each prediction, it is multiplied by a power.

\n
for p in predctions:\n  ensemble += p[i]\nensemble = ensemble/num_of_model\n\nchange to\n\nfor p in predictions:\n  ensemble += p[i]**alpha\nensemble = ensemble/num_of_model\n
\n

hengck23 uses 0.5 for alpha, but 0.5 is not necessarily optimal.
\nFor example, my best submission in RANZCR had an alpha of 0.36.
\nAlpha moves in the range of 0 to 1 in most cases and has one minimum value of loss (or maximum value of AUC).
\nIf you run out of things to do in this competition, you can try exploring Alpha.

\n

Various AUC optimization methods have been posted by @sirishks in the SIIM-ISIC Melanoma Classification competition.

", "messages": [ { "id": 1305300, "postDate": "2021-05-13T07:46:14.460Z", "content": "

This competition has the AUC evaluation metric and ensemble notebook is very powerful.
\nThis reminds me of the RANZCR CLiP - Catheter and Line Position Challenge competition.
\nBy the way, @hengck23 posted a discussion titled one dark magic trick.
\nit means that in the ensemble notes, just before weighting each prediction, it is multiplied by a power.

\n
for p in predctions:\n  ensemble += p[i]\nensemble = ensemble/num_of_model\n\nchange to\n\nfor p in predictions:\n  ensemble += p[i]**alpha\nensemble = ensemble/num_of_model\n
\n

hengck23 uses 0.5 for alpha, but 0.5 is not necessarily optimal.
\nFor example, my best submission in RANZCR had an alpha of 0.36.
\nAlpha moves in the range of 0 to 1 in most cases and has one minimum value of loss (or maximum value of AUC).
\nIf you run out of things to do in this competition, you can try exploring Alpha.

\n

Various AUC optimization methods have been posted by @sirishks in the SIIM-ISIC Melanoma Classification competition.

", "rawMarkdown": "This competition has the AUC evaluation metric and ensemble notebook is very powerful.\nThis reminds me of the [RANZCR CLiP - Catheter and Line Position Challenge](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification) competition.\nBy the way, @hengck23 posted a discussion titled [one dark magic trick](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/211194).\nit means that in the ensemble notes, just before weighting each prediction, it is multiplied by a power.\n```\nfor p in predctions:\n ensemble += p[i]\nensemble = ensemble/num_of_model\n\nchange to\n\nfor p in predictions:\n ensemble += p[i]**alpha\nensemble = ensemble/num_of_model\n```\nhengck23 uses 0.5 for alpha, but 0.5 is not necessarily optimal.\nFor example, my best submission in RANZCR had an alpha of 0.36.\nAlpha moves in the range of 0 to 1 in most cases and has one minimum value of loss (or maximum value of AUC).\nIf you run out of things to do in this competition, you can try exploring Alpha.\n\n[Various AUC optimization methods](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/165653) have been posted by @sirishks in the [SIIM-ISIC Melanoma Classification](https://www.kaggle.com/c/siim-isic-melanoma-classification) competition.", "votes": 39 }, { "id": 1308888, "postDate": "2021-05-15T14:24:19.260Z", "content": "

Thx!
\nThis is my first time participating in the Kaggle competition, so I didn't know this technique using alpha.
\nI have already upvoted.

", "rawMarkdown": "Thx!\nThis is my first time participating in the Kaggle competition, so I didn't know this technique using alpha.\nI have already upvoted.", "votes": 2 }, { "id": 1305328, "postDate": "2021-05-13T08:01:43.563Z", "content": "

In other topic, @ttahara has posted a discussion titled tips: rank averaging in RANZCR.
\nThis method converts the prediction value into a rank, and when ensembling, the average of the ranks is used as the prediction.
\nIn my method this was not the best, but the CV was improved over the usual prediction(usual prediction means 0<pred<1).
\nThe winner of the SIIM-ISIC Melanoma Classification competition, @boliu0, said in the winner's solution that he used this.
\nIt's not just about improving the model, but also post-processing will be the key to win.

", "rawMarkdown": "In other topic, @ttahara has posted a discussion titled [tips: rank averaging](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/205564) in RANZCR.\nThis method converts the prediction value into a rank, and when ensembling, the average of the ranks is used as the prediction.\nIn my method this was not the best, but the CV was improved over the usual prediction(usual prediction means 0Thx!
\nThis is my first time participating in the Kaggle competition, so I didn't know this technique using alpha.
\nI have already upvoted.

", "votes": 2, "replies": [] }, { "id": 1305328, "author_name": "ITK8191", "author_url": "", "post_date": "2021-05-13T08:01:43.563000", "content": "

In other topic, @ttahara has posted a discussion titled tips: rank averaging in RANZCR.
\nThis method converts the prediction value into a rank, and when ensembling, the average of the ranks is used as the prediction.
\nIn my method this was not the best, but the CV was improved over the usual prediction(usual prediction means 0<pred<1).
\nThe winner of the SIIM-ISIC Melanoma Classification competition, @boliu0, said in the winner's solution that he used this.
\nIt's not just about improving the model, but also post-processing will be the key to win.

", "votes": 2, "replies": [] } ], "raw_markdown_by_id": { "1305300": "This competition has the AUC evaluation metric and ensemble notebook is very powerful.\nThis reminds me of the [RANZCR CLiP - Catheter and Line Position Challenge](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification) competition.\nBy the way, @hengck23 posted a discussion titled [one dark magic trick](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/211194).\nit means that in the ensemble notes, just before weighting each prediction, it is multiplied by a power.\n```\nfor p in predctions:\n ensemble += p[i]\nensemble = ensemble/num_of_model\n\nchange to\n\nfor p in predictions:\n ensemble += p[i]**alpha\nensemble = ensemble/num_of_model\n```\nhengck23 uses 0.5 for alpha, but 0.5 is not necessarily optimal.\nFor example, my best submission in RANZCR had an alpha of 0.36.\nAlpha moves in the range of 0 to 1 in most cases and has one minimum value of loss (or maximum value of AUC).\nIf you run out of things to do in this competition, you can try exploring Alpha.\n\n[Various AUC optimization methods](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/165653) have been posted by @sirishks in the [SIIM-ISIC Melanoma Classification](https://www.kaggle.com/c/siim-isic-melanoma-classification) competition.", "1308888": "Thx!\nThis is my first time participating in the Kaggle competition, so I didn't know this technique using alpha.\nI have already upvoted.", "1305328": "In other topic, @ttahara has posted a discussion titled [tips: rank averaging](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/205564) in RANZCR.\nThis method converts the prediction value into a rank, and when ensembling, the average of the ranks is used as the prediction.\nIn my method this was not the best, but the CV was improved over the usual prediction(usual prediction means 0