E.g. photo id 1 is repeated several times for different businesses. How should I understand that?
\n\nE.g. for businesses 3tv9h and 4udt2
", "messages": [ { "id": "106453", "postDate": "01/31/2016 22:28:00", "content": "E.g. photo id 1 is repeated several times for different businesses. How should I understand that?
\n\nE.g. for businesses 3tv9h and 4udt2
", "rawMarkdown": "E.g. photo id 1 is repeated several times for different businesses. How should I understand that? \r\n\r\nE.g. for businesses 3tv9h and 4udt2", "votes": null }, { "id": "106638", "postDate": "02/02/2016 19:07:33", "content": "In the test set there are some businesses which have been added for noise. As mentioned in the data page: 'To deter hand labeling, Kaggle has supplemented the test set with additional "ignored" businesses. These are not counted in the scoring.' See this thread as well: https://www.kaggle.com/c/yelp-restaurant-photo-classification/forums/t/18335/business-ids-in-test.
", "rawMarkdown": "In the test set there are some businesses which have been added for noise. As mentioned in the data page: 'To deter hand labeling, Kaggle has supplemented the test set with additional \"ignored\" businesses. These are not counted in the scoring.' See this thread as well: https://www.kaggle.com/c/yelp-restaurant-photo-classification/forums/t/18335/business-ids-in-test.", "votes": null }, { "id": "109066", "postDate": "02/22/2016 23:48:33", "content": "Apparently they did this a lot. IMHO it was unnecessary.
\n\nWe already have a ridiculous number of images and businesses to process in a competition with a $0 prize... no need to artificially jack up the size of the data and make it more convoluted with fake businesses that break logical algorithms based on the idea that each business has a unique set of photos.
", "rawMarkdown": "Apparently they did this a **lot**. IMHO it was unnecessary. \r\n\r\nWe already have a ridiculous number of images and businesses to process in a competition with a $0 prize... no need to artificially jack up the size of the data and make it more convoluted with fake businesses that break logical algorithms based on the idea that each business has a unique set of photos.", "votes": null }, { "id": "109151", "postDate": "02/23/2016 19:19:40", "content": "There is indeed quite some processing overhead if you process those images multiple times.
\n\nTo avoid that, I do two stages: First scoring of each test image individually and storing the results in a dataframe, and then joining the scores to the businesses in the second stage. This way I don't need to process the images multiple times. Maybe that can work for your approach as well.
", "rawMarkdown": "There is indeed quite some processing overhead if you process those images multiple times. \r\n\r\nTo avoid that, I do two stages: First scoring of each test image individually and storing the results in a dataframe, and then joining the scores to the businesses in the second stage. This way I don't need to process the images multiple times. Maybe that can work for your approach as well.", "votes": null }, { "id": "112917", "postDate": "03/25/2016 02:27:03", "content": "Yeah, this is pervasive :( ..
\n\nThere are 1,190,225 rows in test_photo_to_biz.csv
\n\nThere are only 237,153 files in test_photos
\n\nSo among the 237,152 photos, they're reused on average 5x
\n\nThat probably means that the real business only account for 1/5 of the actual entries and they just sampled from the other businesses randomly to fill in the gaps.
\n\nAttached some screenshots of a table and a graph showing the dupes.
\n\n
In the test set there are some businesses which have been added for noise. As mentioned in the data page: 'To deter hand labeling, Kaggle has supplemented the test set with additional "ignored" businesses. These are not counted in the scoring.' See this thread as well: https://www.kaggle.com/c/yelp-restaurant-photo-classification/forums/t/18335/business-ids-in-test.
", "votes": null, "replies": [] }, { "id": 109066, "author_name": "millerintllc", "author_url": "", "post_date": "02/22/2016 23:48:33", "content": "Apparently they did this a lot. IMHO it was unnecessary.
\n\nWe already have a ridiculous number of images and businesses to process in a competition with a $0 prize... no need to artificially jack up the size of the data and make it more convoluted with fake businesses that break logical algorithms based on the idea that each business has a unique set of photos.
", "votes": null, "replies": [] }, { "id": 109151, "author_name": "allexius", "author_url": "", "post_date": "02/23/2016 19:19:40", "content": "There is indeed quite some processing overhead if you process those images multiple times.
\n\nTo avoid that, I do two stages: First scoring of each test image individually and storing the results in a dataframe, and then joining the scores to the businesses in the second stage. This way I don't need to process the images multiple times. Maybe that can work for your approach as well.
", "votes": null, "replies": [] }, { "id": 112917, "author_name": "frankcarey", "author_url": "", "post_date": "03/25/2016 02:27:03", "content": "Yeah, this is pervasive :( ..
\n\nThere are 1,190,225 rows in test_photo_to_biz.csv
\n\nThere are only 237,153 files in test_photos
\n\nSo among the 237,152 photos, they're reused on average 5x
\n\nThat probably means that the real business only account for 1/5 of the actual entries and they just sampled from the other businesses randomly to fill in the gaps.
\n\nAttached some screenshots of a table and a graph showing the dupes.
\n\n