With the new published notebook that utilizes data leakage in post processing the new best public LB score has dropped from 3.091 to 3.055.
\nIt is crazy to think that only a few weeks ago the top score was in the mid 5’s which is now beaten by the best published notebook. With some impressive deep learning models supported with clever post processing the current first place is near decimals from a sub 3 score.
Do you think more data leakage will result in more lowered scores? What other techniques will help us improve our scores? Take your predictions on how low you think the top score can get (see this post: Estimate of best prediction score) .
\nSee this post on what current first place is saying in their discussion
\n[EDIT] Congrats to mamas & Youri on achieving the first sub 3!
", "messages": [ { "id": "1260328", "postDate": "04/02/2021 02:57:12", "content": "With the new published notebook that utilizes data leakage in post processing the new best public LB score has dropped from 3.091 to 3.055.
\nIt is crazy to think that only a few weeks ago the top score was in the mid 5’s which is now beaten by the best published notebook. With some impressive deep learning models supported with clever post processing the current first place is near decimals from a sub 3 score.
Do you think more data leakage will result in more lowered scores? What other techniques will help us improve our scores? Take your predictions on how low you think the top score can get (see this post: Estimate of best prediction score) .
\nSee this post on what current first place is saying in their discussion
\n[EDIT] Congrats to mamas & Youri on achieving the first sub 3!
", "rawMarkdown": "With the [new published notebook that utilizes data leakage in post processing](https://www.kaggle.com/tomooinubushi/postprocessing-based-on-leakage) the new best public LB score has dropped from 3.091 to 3.055. \nIt is crazy to think that only a few weeks ago the top score was in the mid 5’s which is now beaten by the best published notebook. With some impressive deep learning models supported with clever post processing the current first place is near decimals from a sub 3 score. \n\nDo you think more data leakage will result in more lowered scores? What other techniques will help us improve our scores? Take your predictions on how low you think the top score can get [(see this post: Estimate of best prediction score)](https://www.kaggle.com/c/indoor-location-navigation/discussion/228796) . \n\n[See this post on what current first place is saying in their discussion](https://www.kaggle.com/c/indoor-location-navigation/discussion/230096)\n\n[EDIT] Congrats to mamas & Youri on achieving the first sub 3!", "votes": null }, { "id": "1260394", "postDate": "04/02/2021 04:48:28", "content": "The way I use the leakage is very naive and immature, as you might notice. I do not use any other information including device type, walking speed, and/or missing link in dataset to identify phone users. I could imagine that almost all start and endpoints of test set are deciphered by smart and clever people.
", "rawMarkdown": "The way I use the leakage is very naive and immature, as you might notice. I do not use any other information including device type, walking speed, and/or missing link in dataset to identify phone users. I could imagine that almost all start and endpoints of test set are deciphered by smart and clever people.", "votes": null }, { "id": "1267941", "postDate": "04/09/2021 01:30:59", "content": "Congrats to mamas and Youri on the first sub 3 score(Public LB score of 2.802)
", "rawMarkdown": "Congrats to mamas and Youri on the first sub 3 score(Public LB score of 2.802)", "votes": null } ], "comments": [ { "id": 1260394, "author_name": "tomooinubushi", "author_url": "", "post_date": "04/02/2021 04:48:28", "content": "The way I use the leakage is very naive and immature, as you might notice. I do not use any other information including device type, walking speed, and/or missing link in dataset to identify phone users. I could imagine that almost all start and endpoints of test set are deciphered by smart and clever people.
", "votes": null, "replies": [] }, { "id": 1267941, "author_name": "ravishah1", "author_url": "", "post_date": "04/09/2021 01:30:59", "content": "Congrats to mamas and Youri on the first sub 3 score(Public LB score of 2.802)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1260328": "With the [new published notebook that utilizes data leakage in post processing](https://www.kaggle.com/tomooinubushi/postprocessing-based-on-leakage) the new best public LB score has dropped from 3.091 to 3.055. \nIt is crazy to think that only a few weeks ago the top score was in the mid 5’s which is now beaten by the best published notebook. With some impressive deep learning models supported with clever post processing the current first place is near decimals from a sub 3 score. \n\nDo you think more data leakage will result in more lowered scores? What other techniques will help us improve our scores? Take your predictions on how low you think the top score can get [(see this post: Estimate of best prediction score)](https://www.kaggle.com/c/indoor-location-navigation/discussion/228796) . \n\n[See this post on what current first place is saying in their discussion](https://www.kaggle.com/c/indoor-location-navigation/discussion/230096)\n\n[EDIT] Congrats to mamas & Youri on achieving the first sub 3!", "1260394": "The way I use the leakage is very naive and immature, as you might notice. I do not use any other information including device type, walking speed, and/or missing link in dataset to identify phone users. I could imagine that almost all start and endpoints of test set are deciphered by smart and clever people.", "1267941": "Congrats to mamas and Youri on the first sub 3 score(Public LB score of 2.802)" }, "source": "meta" }