First, thanks to Facebook for organising this competition. It is very different and therefore very interesting!
\n\nUnfortunately, I have not had time to submit a solution but I have enjoyed occasionally reading along to the forums as the competition has progressed. I wanted to share an idea i had for tackling this problem to see if anyone did something similar and whether it worked at all!
\n\nThe biggest difficulty in this competition (in my opinion) was the very large number of classes. This made a lot of great algorithms non-trivial to apply (I suspect this was very intentional of FB to challenge people to work around that). My idea was to change this into a binary classification problem!
\n\nIf we take some kind of join of the training data with itself (probably not the complete join as this would be far too large) and add a column indicating whether the two entries are the same location or not. Now we can run any algorithm for binary classification.
\n\nDid anyone try such an approach? Any success?
", "messages": [ { "id": "126194", "postDate": "07/07/2016 01:24:00", "content": "First, thanks to Facebook for organising this competition. It is very different and therefore very interesting!
\n\nUnfortunately, I have not had time to submit a solution but I have enjoyed occasionally reading along to the forums as the competition has progressed. I wanted to share an idea i had for tackling this problem to see if anyone did something similar and whether it worked at all!
\n\nThe biggest difficulty in this competition (in my opinion) was the very large number of classes. This made a lot of great algorithms non-trivial to apply (I suspect this was very intentional of FB to challenge people to work around that). My idea was to change this into a binary classification problem!
\n\nIf we take some kind of join of the training data with itself (probably not the complete join as this would be far too large) and add a column indicating whether the two entries are the same location or not. Now we can run any algorithm for binary classification.
\n\nDid anyone try such an approach? Any success?
", "rawMarkdown": "First, thanks to Facebook for organising this competition. It is very different and therefore very interesting!\r\n\r\nUnfortunately, I have not had time to submit a solution but I have enjoyed occasionally reading along to the forums as the competition has progressed. I wanted to share an idea i had for tackling this problem to see if anyone did something similar and whether it worked at all!\r\n\r\nThe biggest difficulty in this competition (in my opinion) was the very large number of classes. This made a lot of great algorithms non-trivial to apply (I suspect this was very intentional of FB to challenge people to work around that). My idea was to change this into a binary classification problem!\r\n\r\nIf we take some kind of join of the training data with itself (probably not the complete join as this would be far too large) and add a column indicating whether the two entries are the same location or not. Now we can run any algorithm for binary classification. \r\n\r\nDid anyone try such an approach? Any success?", "votes": null }, { "id": "126296", "postDate": "07/07/2016 14:32:55", "content": "It definitely has its benefits to solve the problem as a binary classification, but the number of classes and data volume obviously makes it quite challenging.
\n\nTom (1st Place) explains his awesome binary classification solution here.
", "rawMarkdown": "It definitely has its benefits to solve the problem as a binary classification, but the number of classes and data volume obviously makes it quite challenging. \r\n\r\nTom (1st Place) explains his awesome binary classification solution [here][1].\r\n\r\n\r\n [1]: https://www.kaggle.com/c/facebook-v-predicting-check-ins/forums/t/22081/1st-place-winning-solution/126233#post126233", "votes": null } ], "comments": [ { "id": 126296, "author_name": "frederik", "author_url": "", "post_date": "07/07/2016 14:32:55", "content": "It definitely has its benefits to solve the problem as a binary classification, but the number of classes and data volume obviously makes it quite challenging.
\n\nTom (1st Place) explains his awesome binary classification solution here.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "126194": "First, thanks to Facebook for organising this competition. It is very different and therefore very interesting!\r\n\r\nUnfortunately, I have not had time to submit a solution but I have enjoyed occasionally reading along to the forums as the competition has progressed. I wanted to share an idea i had for tackling this problem to see if anyone did something similar and whether it worked at all!\r\n\r\nThe biggest difficulty in this competition (in my opinion) was the very large number of classes. This made a lot of great algorithms non-trivial to apply (I suspect this was very intentional of FB to challenge people to work around that). My idea was to change this into a binary classification problem!\r\n\r\nIf we take some kind of join of the training data with itself (probably not the complete join as this would be far too large) and add a column indicating whether the two entries are the same location or not. Now we can run any algorithm for binary classification. \r\n\r\nDid anyone try such an approach? Any success?", "126296": "It definitely has its benefits to solve the problem as a binary classification, but the number of classes and data volume obviously makes it quite challenging. \r\n\r\nTom (1st Place) explains his awesome binary classification solution [here][1].\r\n\r\n\r\n [1]: https://www.kaggle.com/c/facebook-v-predicting-check-ins/forums/t/22081/1st-place-winning-solution/126233#post126233" }, "source": "meta" }