{ "id": 194012, "title": "RSNA STR: Shakeup scatterplot", "url": "/competitions/rsna-str-pulmonary-embolism-detection/discussion/194012", "author_name": "Carl McBride Ellis", "post_date": "2020-10-30T07:58:13.554000", "votes": 11, "comment_count": 0, "views": 0, "content": "

For the curious, the RSNA STR Pulmonary Embolism Detection data is now available in the meta kaggle dataset, so here is a snapshot of the shakeup for this competition. For a fully interactive version one can use the shakeup interactive scatterplot maker in conjunction with search_string = \"RSNA STR\".

\n

\"\"

\n

One can see that in general this was a fairly 'well behaved' shakeup, with an interesting box in the middle comprised of five layers. This particular box is actually an isoscore region, and I suggest that is the result of people all submitting the MONAI 3D CNN - Inference notebook written by @boliu0, which was forked more than 580 times and whose Public score was 0.295 and Private score was 0.309. The stratification may be due to something as small as rounding errors in computing the log loss score deciding the minutia of the shakeup in this region. Indeed, a close inspection of the meta kaggle Submissions.csv file shows that the PublicScoreFullPrecision of 0.295715 resolves into a PrivateScoreFullPrecision of either 0.30902, 0.309023 or 0.309026 with no obviously discernible pattern, and accounts for three of the five layers.

\n

All the best,
\ncarl

", "messages": [ { "id": 1064519, "postDate": "2020-10-30T07:58:13.553Z", "content": "

For the curious, the RSNA STR Pulmonary Embolism Detection data is now available in the meta kaggle dataset, so here is a snapshot of the shakeup for this competition. For a fully interactive version one can use the shakeup interactive scatterplot maker in conjunction with search_string = \"RSNA STR\".

\n

\"\"

\n

One can see that in general this was a fairly 'well behaved' shakeup, with an interesting box in the middle comprised of five layers. This particular box is actually an isoscore region, and I suggest that is the result of people all submitting the MONAI 3D CNN - Inference notebook written by @boliu0, which was forked more than 580 times and whose Public score was 0.295 and Private score was 0.309. The stratification may be due to something as small as rounding errors in computing the log loss score deciding the minutia of the shakeup in this region. Indeed, a close inspection of the meta kaggle Submissions.csv file shows that the PublicScoreFullPrecision of 0.295715 resolves into a PrivateScoreFullPrecision of either 0.30902, 0.309023 or 0.309026 with no obviously discernible pattern, and accounts for three of the five layers.

\n

All the best,
\ncarl

", "rawMarkdown": "For the curious, the RSNA STR Pulmonary Embolism Detection data is now available in the [meta kaggle dataset](https://www.kaggle.com/kaggle/meta-kaggle), so here is a snapshot of the shakeup for this competition. For a fully interactive version one can use the [shakeup interactive scatterplot maker](https://www.kaggle.com/carlmcbrideellis/shakeup-interactive-scatterplot-maker) in conjunction with `search_string = \"RSNA STR\"`.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4051350%2Ff6968fabe3e359d65774c7e7d93c1d98%2FRSNA_STR_scatterPlot.png?generation=1604044488178601&alt=media)\n\nOne can see that in general this was a fairly 'well behaved' shakeup, with an interesting box in the middle comprised of five layers. This particular box is actually an [isoscore region](https://www.kaggle.com/carlmcbrideellis/shakeup-scatterplots-boxes-strings-and-things), and I suggest that is the result of people all submitting the [MONAI 3D CNN - Inference](https://www.kaggle.com/boliu0/monai-3d-cnn-inference) notebook written by @boliu0, which was forked more than 580 times and whose Public score was `0.295` and Private score was `0.309`. The stratification may be due to something as small as rounding errors in computing the log loss score deciding the minutia of the shakeup in this region. Indeed, a close inspection of the meta kaggle `Submissions.csv` file shows that the `PublicScoreFullPrecision` of `0.295715` resolves into a `PrivateScoreFullPrecision` of either `0.30902`, `0.309023` or `0.309026` with no obviously discernible pattern, and accounts for three of the five layers.\n\nAll the best,\ncarl ", "votes": 11 } ], "comments": [], "raw_markdown_by_id": { "1064519": "For the curious, the RSNA STR Pulmonary Embolism Detection data is now available in the [meta kaggle dataset](https://www.kaggle.com/kaggle/meta-kaggle), so here is a snapshot of the shakeup for this competition. For a fully interactive version one can use the [shakeup interactive scatterplot maker](https://www.kaggle.com/carlmcbrideellis/shakeup-interactive-scatterplot-maker) in conjunction with `search_string = \"RSNA STR\"`.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4051350%2Ff6968fabe3e359d65774c7e7d93c1d98%2FRSNA_STR_scatterPlot.png?generation=1604044488178601&alt=media)\n\nOne can see that in general this was a fairly 'well behaved' shakeup, with an interesting box in the middle comprised of five layers. This particular box is actually an [isoscore region](https://www.kaggle.com/carlmcbrideellis/shakeup-scatterplots-boxes-strings-and-things), and I suggest that is the result of people all submitting the [MONAI 3D CNN - Inference](https://www.kaggle.com/boliu0/monai-3d-cnn-inference) notebook written by @boliu0, which was forked more than 580 times and whose Public score was `0.295` and Private score was `0.309`. The stratification may be due to something as small as rounding errors in computing the log loss score deciding the minutia of the shakeup in this region. Indeed, a close inspection of the meta kaggle `Submissions.csv` file shows that the `PublicScoreFullPrecision` of `0.295715` resolves into a `PrivateScoreFullPrecision` of either `0.30902`, `0.309023` or `0.309026` with no obviously discernible pattern, and accounts for three of the five layers.\n\nAll the best,\ncarl " } }