Hi everyone!
\n\nI'm new here and to machine learning and I'm just starting off.So i would really appreciate some guidance on how i can get started with this and some ideas on how to approach this.Some online resources on how to tackle this would also be great!Thanks in advance and all the best to you all!
\n\nThanks\nBalaji
", "messages": [ { "id": "106917", "postDate": "02/04/2016 21:58:18", "content": "Hi everyone!
\n\nI'm new here and to machine learning and I'm just starting off.So i would really appreciate some guidance on how i can get started with this and some ideas on how to approach this.Some online resources on how to tackle this would also be great!Thanks in advance and all the best to you all!
\n\nThanks\nBalaji
", "rawMarkdown": "Hi everyone!\r\n\r\nI'm new here and to machine learning and I'm just starting off.So i would really appreciate some guidance on how i can get started with this and some ideas on how to approach this.Some online resources on how to tackle this would also be great!Thanks in advance and all the best to you all!\r\n\r\nThanks\r\nBalaji", "votes": null }, { "id": "107016", "postDate": "02/05/2016 21:44:07", "content": "I am in the same situation as you are. I very new to machine learning, but I have done some reading the past few days on how to approach this problem, so I will share what I have found.
\n\nTo start I have been looking at multi-label classification (https://en.wikipedia.org/wiki/Multi-label_classification) as that seems to be the type of problem we are dealing with here. Wikipedia mentions different ML techniques that can be used to solve these types of problems.
", "rawMarkdown": "I am in the same situation as you are. I very new to machine learning, but I have done some reading the past few days on how to approach this problem, so I will share what I have found. \r\n\r\nTo start I have been looking at multi-label classification (https://en.wikipedia.org/wiki/Multi-label_classification) as that seems to be the type of problem we are dealing with here. Wikipedia mentions different ML techniques that can be used to solve these types of problems.", "votes": null }, { "id": "107031", "postDate": "02/05/2016 23:51:13", "content": "Hi,
\n\nThank you so much for all that info!Highly useful and I will definitely look into that!
", "rawMarkdown": "Hi,\r\n\r\nThank you so much for all that info!Highly useful and I will definitely look into that!", "votes": null }, { "id": "107033", "postDate": "02/06/2016 00:17:42", "content": "After putting more thought into, I think perhaps my original idea of turning it into a multi-class classification problem is probably not the best thing to do. When looking at the data I see that there are a total of only 173 combinations of labels. That means that there wouldn't be training data for 339 of the possible classes. I my guess would be if that approach is taken the model wouldn't generalize as well.
", "rawMarkdown": "After putting more thought into, I think perhaps my original idea of turning it into a multi-class classification problem is probably not the best thing to do. When looking at the data I see that there are a total of only 173 combinations of labels. That means that there wouldn't be training data for 339 of the possible classes. I my guess would be if that approach is taken the model wouldn't generalize as well.", "votes": null }, { "id": "107164", "postDate": "02/07/2016 22:46:22", "content": "I have also started learning ML through this problem. Particularly this resource http://cs231n.github.io/ seems to be helpful in developing an understanding of different techniques that can be used.
", "rawMarkdown": "I have also started learning ML through this problem. Particularly this resource http://cs231n.github.io/ seems to be helpful in developing an understanding of different techniques that can be used.", "votes": null }, { "id": "108431", "postDate": "02/17/2016 08:47:43", "content": "Hi Guys,
\n\nIf I turn this problem into Multi-Label Classification problem and train the images based on their labels i.e business attributes. Then the output that I would get is a binary vector for the test images containing the labels. But I am still not able to figure how to map the biz ids in test set to train set. How to figure out which are the valid biz id for a image in the test set because in test set same image has multiple biz ids.
\n\nPlease suggest.
", "rawMarkdown": "Hi Guys,\r\n\r\nIf I turn this problem into Multi-Label Classification problem and train the images based on their labels i.e business attributes. Then the output that I would get is a binary vector for the test images containing the labels. But I am still not able to figure how to map the biz ids in test set to train set. How to figure out which are the valid biz id for a image in the test set because in test set same image has multiple biz ids. \r\n\r\nPlease suggest.", "votes": null }, { "id": "108466", "postDate": "02/17/2016 15:28:38", "content": "my thinking currently is as follows:
\n\nWhat do you think?\nJohannes
\n\nEdit:\nAlternatively combine steps 3 and 4 above.\nTake output metrics from image classification directly into business classifier without collapsing them down, the downside is that it doesn't take into account the individual images' classification confidences. i.e.
\n\nbusiness_id, 0_min, 0_max, 0_majority, 0_deciles..., 0_sd, ... 8_sd, num_images\n[quote=Johannes Ahlmann;108466]
\n\nmy thinking currently is as follows:
\n\nWhat do you think?\n[/quote]
\n\nYou are on the right track - that's more or less my current approach, with some additional tweaks to optimize for the evaluation metric. Number of images per business as a feature didn't help in my setup, but it's worth to look for any kind of meta-data, apart from the image raw data.
", "rawMarkdown": "[quote=Johannes Ahlmann;108466]\r\n\r\nmy thinking currently is as follows:\r\n\r\n- assign all images of a business the business categories\r\n- train 9 binary classifiers on the images\r\n- combine image classifications for a business (majority vote, take classification confidence into account, train a seprate classifier based on min/max/mean/median/stddev/centile range against business classes, etc.)\r\n- train classifier on businesses with combined image classification (class 0-8) plus additional features (i.e. number of images per business)\r\n\r\nWhat do you think?\r\n[/quote]\r\n\r\nYou are on the right track - that's more or less my current approach, with some additional tweaks to optimize for the evaluation metric. Number of images per business as a feature didn't help in my setup, but it's worth to look for any kind of meta-data, apart from the image raw data.", "votes": null } ], "comments": [ { "id": 107016, "author_name": "djatha", "author_url": "", "post_date": "02/05/2016 21:44:07", "content": "I am in the same situation as you are. I very new to machine learning, but I have done some reading the past few days on how to approach this problem, so I will share what I have found.
\n\nTo start I have been looking at multi-label classification (https://en.wikipedia.org/wiki/Multi-label_classification) as that seems to be the type of problem we are dealing with here. Wikipedia mentions different ML techniques that can be used to solve these types of problems.
", "votes": null, "replies": [] }, { "id": 107031, "author_name": "balag59", "author_url": "", "post_date": "02/05/2016 23:51:13", "content": "Hi,
\n\nThank you so much for all that info!Highly useful and I will definitely look into that!
", "votes": null, "replies": [] }, { "id": 107033, "author_name": "djatha", "author_url": "", "post_date": "02/06/2016 00:17:42", "content": "After putting more thought into, I think perhaps my original idea of turning it into a multi-class classification problem is probably not the best thing to do. When looking at the data I see that there are a total of only 173 combinations of labels. That means that there wouldn't be training data for 339 of the possible classes. I my guess would be if that approach is taken the model wouldn't generalize as well.
", "votes": null, "replies": [] }, { "id": 107164, "author_name": "nav13n", "author_url": "", "post_date": "02/07/2016 22:46:22", "content": "I have also started learning ML through this problem. Particularly this resource http://cs231n.github.io/ seems to be helpful in developing an understanding of different techniques that can be used.
", "votes": null, "replies": [] }, { "id": 108431, "author_name": "rayush7", "author_url": "", "post_date": "02/17/2016 08:47:43", "content": "Hi Guys,
\n\nIf I turn this problem into Multi-Label Classification problem and train the images based on their labels i.e business attributes. Then the output that I would get is a binary vector for the test images containing the labels. But I am still not able to figure how to map the biz ids in test set to train set. How to figure out which are the valid biz id for a image in the test set because in test set same image has multiple biz ids.
\n\nPlease suggest.
", "votes": null, "replies": [] }, { "id": 108466, "author_name": "ahlmann", "author_url": "", "post_date": "02/17/2016 15:28:38", "content": "my thinking currently is as follows:
\n\nWhat do you think?\nJohannes
\n\nEdit:\nAlternatively combine steps 3 and 4 above.\nTake output metrics from image classification directly into business classifier without collapsing them down, the downside is that it doesn't take into account the individual images' classification confidences. i.e.
\n\nbusiness_id, 0_min, 0_max, 0_majority, 0_deciles..., 0_sd, ... 8_sd, num_images\n[quote=Johannes Ahlmann;108466]
\n\nmy thinking currently is as follows:
\n\nWhat do you think?\n[/quote]
\n\nYou are on the right track - that's more or less my current approach, with some additional tweaks to optimize for the evaluation metric. Number of images per business as a feature didn't help in my setup, but it's worth to look for any kind of meta-data, apart from the image raw data.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "106917": "Hi everyone!\r\n\r\nI'm new here and to machine learning and I'm just starting off.So i would really appreciate some guidance on how i can get started with this and some ideas on how to approach this.Some online resources on how to tackle this would also be great!Thanks in advance and all the best to you all!\r\n\r\nThanks\r\nBalaji", "107016": "I am in the same situation as you are. I very new to machine learning, but I have done some reading the past few days on how to approach this problem, so I will share what I have found. \r\n\r\nTo start I have been looking at multi-label classification (https://en.wikipedia.org/wiki/Multi-label_classification) as that seems to be the type of problem we are dealing with here. Wikipedia mentions different ML techniques that can be used to solve these types of problems.", "107031": "Hi,\r\n\r\nThank you so much for all that info!Highly useful and I will definitely look into that!", "107033": "After putting more thought into, I think perhaps my original idea of turning it into a multi-class classification problem is probably not the best thing to do. When looking at the data I see that there are a total of only 173 combinations of labels. That means that there wouldn't be training data for 339 of the possible classes. I my guess would be if that approach is taken the model wouldn't generalize as well.", "107164": "I have also started learning ML through this problem. Particularly this resource http://cs231n.github.io/ seems to be helpful in developing an understanding of different techniques that can be used.", "108431": "Hi Guys,\r\n\r\nIf I turn this problem into Multi-Label Classification problem and train the images based on their labels i.e business attributes. Then the output that I would get is a binary vector for the test images containing the labels. But I am still not able to figure how to map the biz ids in test set to train set. How to figure out which are the valid biz id for a image in the test set because in test set same image has multiple biz ids. \r\n\r\nPlease suggest.", "108466": "my thinking currently is as follows:\r\n\r\n- assign all images of a business the business categories\r\n- train 9 binary classifiers on the images\r\n- combine image classifications for a business (majority vote, take classification confidence into account, train a seprate classifier based on min/max/mean/median/stddev/centile range against business classes, etc.)\r\n- train classifier on businesses with combined image classification (class 0-8) plus additional features (i.e. number of images per business)\r\n\r\nWhat do you think?\r\nJohannes\r\n\r\n**Edit**:\r\nAlternatively combine steps 3 and 4 above.\r\nTake output metrics from image classification directly into business classifier without collapsing them down, the downside is that it doesn't take into account the individual images' classification confidences. i.e.\r\n\r\n business_id, 0_min, 0_max, 0_majority, 0_deciles..., 0_sd, ... 8_sd, num_images", "108748": "[quote=Johannes Ahlmann;108466]\r\n\r\nmy thinking currently is as follows:\r\n\r\n- assign all images of a business the business categories\r\n- train 9 binary classifiers on the images\r\n- combine image classifications for a business (majority vote, take classification confidence into account, train a seprate classifier based on min/max/mean/median/stddev/centile range against business classes, etc.)\r\n- train classifier on businesses with combined image classification (class 0-8) plus additional features (i.e. number of images per business)\r\n\r\nWhat do you think?\r\n[/quote]\r\n\r\nYou are on the right track - that's more or less my current approach, with some additional tweaks to optimize for the evaluation metric. Number of images per business as a feature didn't help in my setup, but it's worth to look for any kind of meta-data, apart from the image raw data." }, "source": "meta" }