For most labels, I'm able to get decent predictions (~80%). After training, I expect to see something like this graph where most of the errors happen when the prediction is halfway between the label not existing (0) and existing(1).
\n\nSee Screen Shot 2016-04-10 at 8.58.31 PM.png
\n\nHowever, what I end up getting (just for category 3) is that the errors seem to be spread out evenly. I've tried a few different models to get at it, but they all only get somewhere in the 50% - 60% range. Also outside seating for most restaurants is usually a pretty rare thing, yet it seems to show up around 50% of the time in the training set, which strikes me as odd.
\n\nSee Screen Shot 2016-04-10 at 8.58.41 PM.png
\n\nAnyone have a clue here, or getting much better results?
", "messages": [ { "id": "114466", "postDate": "04/11/2016 05:49:36", "content": "For most labels, I'm able to get decent predictions (~80%). After training, I expect to see something like this graph where most of the errors happen when the prediction is halfway between the label not existing (0) and existing(1).
\n\nSee Screen Shot 2016-04-10 at 8.58.31 PM.png
\n\nHowever, what I end up getting (just for category 3) is that the errors seem to be spread out evenly. I've tried a few different models to get at it, but they all only get somewhere in the 50% - 60% range. Also outside seating for most restaurants is usually a pretty rare thing, yet it seems to show up around 50% of the time in the training set, which strikes me as odd.
\n\nSee Screen Shot 2016-04-10 at 8.58.41 PM.png
\n\nAnyone have a clue here, or getting much better results?
", "rawMarkdown": "For most labels, I'm able to get decent predictions (~80%). After training, I expect to see something like this graph where most of the errors happen when the prediction is halfway between the label not existing (0) and existing(1). \r\n\r\nSee Screen Shot 2016-04-10 at 8.58.31 PM.png\r\n\r\nHowever, what I end up getting (just for category 3) is that the errors seem to be spread out evenly. I've tried a few different models to get at it, but they all only get somewhere in the 50% - 60% range. Also outside seating for most restaurants is usually a pretty rare thing, yet it seems to show up around 50% of the time in the training set, which strikes me as odd.\r\n\r\nSee Screen Shot 2016-04-10 at 8.58.41 PM.png\r\n\r\nAnyone have a clue here, or getting much better results?", "votes": null }, { "id": "114678", "postDate": "04/12/2016 20:47:30", "content": "I have the same problems with this category. My best model gets following scores:
\n\n[ 0.77497864 0.85935432 0.89804971 0.73494768 0.83317268 0.9062289\n 0.95272082 0.80270159 0.91307145]
", "rawMarkdown": "I have the same problems with this category. My best model gets following scores:\r\n \r\n[ 0.77497864 0.85935432 0.89804971 0.73494768 0.83317268 0.9062289\r\n 0.95272082 0.80270159 0.91307145]", "votes": null }, { "id": "114687", "postDate": "04/12/2016 21:35:21", "content": "It seems that there are too few outdoor photos.
", "rawMarkdown": "It seems that there are too few outdoor photos.", "votes": null }, { "id": "114852", "postDate": "04/14/2016 10:25:51", "content": "Yes there was definitely something wrong with the 3rd category , during the competition when i was in the middle of applying MISVM on each label. All labels except the 3rd got around 80 % accuracy , while it was suffering on 56-60 % accuracy.
", "rawMarkdown": "Yes there was definitely something wrong with the 3rd category , during the competition when i was in the middle of applying MISVM on each label. All labels except the 3rd got around 80 % accuracy , while it was suffering on 56-60 % accuracy.", "votes": null } ], "comments": [ { "id": 114678, "author_name": "u1234x1234", "author_url": "", "post_date": "04/12/2016 20:47:30", "content": "I have the same problems with this category. My best model gets following scores:
\n\n[ 0.77497864 0.85935432 0.89804971 0.73494768 0.83317268 0.9062289\n 0.95272082 0.80270159 0.91307145]
", "votes": null, "replies": [] }, { "id": 114687, "author_name": "pinocchio", "author_url": "", "post_date": "04/12/2016 21:35:21", "content": "It seems that there are too few outdoor photos.
", "votes": null, "replies": [] }, { "id": 114852, "author_name": "yardstick17", "author_url": "", "post_date": "04/14/2016 10:25:51", "content": "Yes there was definitely something wrong with the 3rd category , during the competition when i was in the middle of applying MISVM on each label. All labels except the 3rd got around 80 % accuracy , while it was suffering on 56-60 % accuracy.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "114466": "For most labels, I'm able to get decent predictions (~80%). After training, I expect to see something like this graph where most of the errors happen when the prediction is halfway between the label not existing (0) and existing(1). \r\n\r\nSee Screen Shot 2016-04-10 at 8.58.31 PM.png\r\n\r\nHowever, what I end up getting (just for category 3) is that the errors seem to be spread out evenly. I've tried a few different models to get at it, but they all only get somewhere in the 50% - 60% range. Also outside seating for most restaurants is usually a pretty rare thing, yet it seems to show up around 50% of the time in the training set, which strikes me as odd.\r\n\r\nSee Screen Shot 2016-04-10 at 8.58.41 PM.png\r\n\r\nAnyone have a clue here, or getting much better results?", "114678": "I have the same problems with this category. My best model gets following scores:\r\n \r\n[ 0.77497864 0.85935432 0.89804971 0.73494768 0.83317268 0.9062289\r\n 0.95272082 0.80270159 0.91307145]", "114687": "It seems that there are too few outdoor photos.", "114852": "Yes there was definitely something wrong with the 3rd category , during the competition when i was in the middle of applying MISVM on each label. All labels except the 3rd got around 80 % accuracy , while it was suffering on 56-60 % accuracy." }, "source": "meta" }