As of now, I am planning to use the floors generated in the 99% Accurate Floor Model notebook. I recently realized how important floors really are to our score. If everyone uses the same floors from that notebook, there will not be much impact, but if someone can have better floors on the private LB, that could put them at the top of the private LB.
\nTo understand the importance of floors on our score let's consider the evaluation metric (linked). The floor penalty is set at 15. What does that mean? Well let’s consider the floor addition of the evaluation alone. This can be calculated with (15 x incorrect predictions)/total predictions (and this suggests you are only ever off by 1 floor) . If we have 99% accurate floors, that would be (15x1)/100 == .15 added to our x,y error. In comparison, if we have 90% accurate floors, that would be (15x10)/100 == 1.5 added to our x,y error. This goes to show how fast incorrect floors can raise your error.
\nSince many of us are using floors as a feature in our models (as I explain in my previous post) and in our post processing such as snap to grid, floors impact all 3 elements of our submission (x, y, and floor).
\nIf someone has somehow managed to generate floor predictions better than the 99% model in private LB they could rise to the top of the private LB.
\nWhat floor predictions are you using? Are you using floors as a feature and in your post processing?
", "messages": [ { "id": "1276020", "postDate": "04/17/2021 02:10:48", "content": "As of now, I am planning to use the floors generated in the 99% Accurate Floor Model notebook. I recently realized how important floors really are to our score. If everyone uses the same floors from that notebook, there will not be much impact, but if someone can have better floors on the private LB, that could put them at the top of the private LB.
\nTo understand the importance of floors on our score let's consider the evaluation metric (linked). The floor penalty is set at 15. What does that mean? Well let’s consider the floor addition of the evaluation alone. This can be calculated with (15 x incorrect predictions)/total predictions (and this suggests you are only ever off by 1 floor) . If we have 99% accurate floors, that would be (15x1)/100 == .15 added to our x,y error. In comparison, if we have 90% accurate floors, that would be (15x10)/100 == 1.5 added to our x,y error. This goes to show how fast incorrect floors can raise your error.
\nSince many of us are using floors as a feature in our models (as I explain in my previous post) and in our post processing such as snap to grid, floors impact all 3 elements of our submission (x, y, and floor).
\nIf someone has somehow managed to generate floor predictions better than the 99% model in private LB they could rise to the top of the private LB.
\nWhat floor predictions are you using? Are you using floors as a feature and in your post processing?
", "rawMarkdown": "As of now, I am planning to use the floors generated in the [99% Accurate Floor Model notebook](https://www.kaggle.com/nigelhenry/simple-99-accurate-floor-model). I recently realized how important floors really are to our score. If everyone uses the same floors from that notebook, there will not be much impact, but if someone can have better floors on the private LB, that could put them at the top of the private LB.\n\nTo understand the importance of floors on our score let's consider the [evaluation metric (linked)](https://www.kaggle.com/c/indoor-location-navigation/overview/evaluation). The floor penalty is set at 15. What does that mean? Well let’s consider the floor addition of the evaluation alone. This can be calculated with (15 x incorrect predictions)/total predictions (and this suggests you are only ever off by 1 floor) . If we have 99% accurate floors, that would be (15x1)/100 == .15 added to our x,y error. In comparison, if we have 90% accurate floors, that would be (15x10)/100 == 1.5 added to our x,y error. This goes to show how fast incorrect floors can raise your error.\n\nSince many of us are using [floors as a feature in our models (as I explain in my previous post) ](https://www.kaggle.com/c/indoor-location-navigation/discussion/228887)and in our post processing such as [snap to grid](https://www.kaggle.com/robikscube/indoor-navigation-snap-to-grid-post-processing), floors impact all 3 elements of our submission (x, y, and floor).\n\nIf someone has somehow managed to generate floor predictions better than the 99% model in private LB they could rise to the top of the private LB.\n\nWhat floor predictions are you using? Are you using floors as a feature and in your post processing?", "votes": null }, { "id": "1276032", "postDate": "04/17/2021 02:41:10", "content": "I totally agree with you.
\nI was thinking that we should put bit more effort on floor prediction.
Nice Observation buddy..
", "rawMarkdown": "Nice Observation buddy..", "votes": null }, { "id": "1276427", "postDate": "04/17/2021 14:49:04", "content": "I think I have more accurate floors than the 99% accurate. I'm using floors as an input, but if I correct them with my \"better\" floors before post-processing, I get a better score. I need to go back and update my scripts to use my better floors.
\nFloor prediction is important but there's something else needed to get sub 4.
", "rawMarkdown": "I think I have more accurate floors than the 99% accurate. I'm using floors as an input, but if I correct them with my \"better\" floors before post-processing, I get a better score. I need to go back and update my scripts to use my better floors. \n\nFloor prediction is important but there's something else needed to get sub 4.", "votes": null }, { "id": "1276545", "postDate": "04/17/2021 17:21:15", "content": "Totally agree
", "rawMarkdown": "Totally agree", "votes": null }, { "id": "1276653", "postDate": "04/17/2021 19:47:12", "content": "Wow, if you have better floors than the 99% notebook and use them as a feature, you could do really well on the private LB. I agree there is something else needed to get sub 4.
", "rawMarkdown": "Wow, if you have better floors than the 99% notebook and use them as a feature, you could do really well on the private LB. I agree there is something else needed to get sub 4.", "votes": null }, { "id": "1276654", "postDate": "04/17/2021 19:47:50", "content": "Yeah, I haven’t really worked on anything floor specific since the 99% floor notebook, but I’m starting to put some effort into floor prediction now.
", "rawMarkdown": "Yeah, I haven’t really worked on anything floor specific since the 99% floor notebook, but I’m starting to put some effort into floor prediction now.", "votes": null } ], "comments": [ { "id": 1276032, "author_name": "kokitanisaka", "author_url": "", "post_date": "04/17/2021 02:41:10", "content": "I totally agree with you.
\nI was thinking that we should put bit more effort on floor prediction.
Yeah, I haven’t really worked on anything floor specific since the 99% floor notebook, but I’m starting to put some effort into floor prediction now.
", "votes": null, "replies": [] } ] }, { "id": 1276405, "author_name": "abhishek252", "author_url": "", "post_date": "04/17/2021 14:27:22", "content": "Nice Observation buddy..
", "votes": null, "replies": [] }, { "id": 1276427, "author_name": "ericfreeman", "author_url": "", "post_date": "04/17/2021 14:49:04", "content": "I think I have more accurate floors than the 99% accurate. I'm using floors as an input, but if I correct them with my \"better\" floors before post-processing, I get a better score. I need to go back and update my scripts to use my better floors.
\nFloor prediction is important but there's something else needed to get sub 4.
", "votes": null, "replies": [ { "id": 1276653, "author_name": "ravishah1", "author_url": "", "post_date": "04/17/2021 19:47:12", "content": "Wow, if you have better floors than the 99% notebook and use them as a feature, you could do really well on the private LB. I agree there is something else needed to get sub 4.
", "votes": null, "replies": [] } ] }, { "id": 1276545, "author_name": "enric1296", "author_url": "", "post_date": "04/17/2021 17:21:15", "content": "Totally agree
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1276020": "As of now, I am planning to use the floors generated in the [99% Accurate Floor Model notebook](https://www.kaggle.com/nigelhenry/simple-99-accurate-floor-model). I recently realized how important floors really are to our score. If everyone uses the same floors from that notebook, there will not be much impact, but if someone can have better floors on the private LB, that could put them at the top of the private LB.\n\nTo understand the importance of floors on our score let's consider the [evaluation metric (linked)](https://www.kaggle.com/c/indoor-location-navigation/overview/evaluation). The floor penalty is set at 15. What does that mean? Well let’s consider the floor addition of the evaluation alone. This can be calculated with (15 x incorrect predictions)/total predictions (and this suggests you are only ever off by 1 floor) . If we have 99% accurate floors, that would be (15x1)/100 == .15 added to our x,y error. In comparison, if we have 90% accurate floors, that would be (15x10)/100 == 1.5 added to our x,y error. This goes to show how fast incorrect floors can raise your error.\n\nSince many of us are using [floors as a feature in our models (as I explain in my previous post) ](https://www.kaggle.com/c/indoor-location-navigation/discussion/228887)and in our post processing such as [snap to grid](https://www.kaggle.com/robikscube/indoor-navigation-snap-to-grid-post-processing), floors impact all 3 elements of our submission (x, y, and floor).\n\nIf someone has somehow managed to generate floor predictions better than the 99% model in private LB they could rise to the top of the private LB.\n\nWhat floor predictions are you using? Are you using floors as a feature and in your post processing?", "1276032": "I totally agree with you. \nI was thinking that we should put bit more effort on floor prediction.", "1276405": "Nice Observation buddy..", "1276427": "I think I have more accurate floors than the 99% accurate. I'm using floors as an input, but if I correct them with my \"better\" floors before post-processing, I get a better score. I need to go back and update my scripts to use my better floors. \n\nFloor prediction is important but there's something else needed to get sub 4.", "1276545": "Totally agree", "1276653": "Wow, if you have better floors than the 99% notebook and use them as a feature, you could do really well on the private LB. I agree there is something else needed to get sub 4.", "1276654": "Yeah, I haven’t really worked on anything floor specific since the 99% floor notebook, but I’m starting to put some effort into floor prediction now." }, "source": "meta" }