\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1035002%2F84308f9c25c81b056588683a72d534a7%2Fall_points.png?generation=1605642213442813&alt=media\")
One of the things I started toying with was visualizing the error per intersection/road to see if potentially some specific area is undertrained or particularly problematic.
\nHere is a plot of all of the centroids for the validation set
\n
All of the validation points with points scaled by error^(1/1.5)
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1035002%2F341124997870d8a4ed6f80a5114cde01%2Fscaled_all_points.png?generation=1605642308639494&alt=media\")
And here there is all of the points with over 1k error with (1/2) scaling
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1035002%2F979f8cce01ed72cf5c0bb51f721ad97f%2Ftop_scaled.png?generation=1605642434710487&alt=media\")
Hard to derive any great insight from these plots, but still interesting to look at. Maybe there is some method to additionally train specifically around the areas where high error occurs and blunt some of those extreme error values
", "messages": [ { "id": "1082323", "postDate": "11/17/2020 19:49:31", "content": "One of the things I started toying with was visualizing the error per intersection/road to see if potentially some specific area is undertrained or particularly problematic.
\nHere is a plot of all of the centroids for the validation set
\n
All of the validation points with points scaled by error^(1/1.5)
\n
And here there is all of the points with over 1k error with (1/2) scaling
\n
Hard to derive any great insight from these plots, but still interesting to look at. Maybe there is some method to additionally train specifically around the areas where high error occurs and blunt some of those extreme error values
", "rawMarkdown": "One of the things I started toying with was visualizing the error per intersection/road to see if potentially some specific area is undertrained or particularly problematic. \n\nHere is a plot of all of the centroids for the validation set\n\n\nAll of the validation points with points scaled by error^(1/1.5)\n\n\n\nAnd here there is all of the points with over 1k error with (1/2) scaling\n\n\n\n\nHard to derive any great insight from these plots, but still interesting to look at. Maybe there is some method to additionally train specifically around the areas where high error occurs and blunt some of those extreme error values", "votes": null }, { "id": "1082404", "postDate": "11/17/2020 21:44:36", "content": "What about something like this :) WARNING: outdated data and non-representative numbers.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F257647%2Fe7b340c2b2a537935c5f51533387aa46%2Findex.png?generation=1605649422587573&alt=media\")
Very interesting. How did you form those segments? Just via clustering?
", "rawMarkdown": "Very interesting. How did you form those segments? Just via clustering?", "votes": null }, { "id": "1082430", "postDate": "11/17/2020 22:09:44", "content": "Yes, k-means based on centroid.
", "rawMarkdown": "Yes, k-means based on centroid.", "votes": null }, { "id": "1082777", "postDate": "11/18/2020 08:10:34", "content": "error is related to the velocity of the agent I think. it may make more sense to build a predictor for difference cluster
", "rawMarkdown": "error is related to the velocity of the agent I think. it may make more sense to build a predictor for difference cluster", "votes": null } ], "comments": [ { "id": 1082404, "author_name": "ggrizzly", "author_url": "", "post_date": "11/17/2020 21:44:36", "content": "What about something like this :) WARNING: outdated data and non-representative numbers.
\n
Very interesting. How did you form those segments? Just via clustering?
", "votes": null, "replies": [] }, { "id": 1082430, "author_name": "ggrizzly", "author_url": "", "post_date": "11/17/2020 22:09:44", "content": "Yes, k-means based on centroid.
", "votes": null, "replies": [] }, { "id": 1082777, "author_name": "hengck23", "author_url": "", "post_date": "11/18/2020 08:10:34", "content": "error is related to the velocity of the agent I think. it may make more sense to build a predictor for difference cluster
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1082323": "One of the things I started toying with was visualizing the error per intersection/road to see if potentially some specific area is undertrained or particularly problematic. \n\nHere is a plot of all of the centroids for the validation set\n\n\nAll of the validation points with points scaled by error^(1/1.5)\n\n\n\nAnd here there is all of the points with over 1k error with (1/2) scaling\n\n\n\n\nHard to derive any great insight from these plots, but still interesting to look at. Maybe there is some method to additionally train specifically around the areas where high error occurs and blunt some of those extreme error values", "1082404": "What about something like this :) WARNING: outdated data and non-representative numbers.\n", "1082420": "Very interesting. How did you form those segments? Just via clustering?", "1082430": "Yes, k-means based on centroid.", "1082777": "error is related to the velocity of the agent I think. it may make more sense to build a predictor for difference cluster" }, "source": "meta" }