I still have not seen an adequate explanation of how deal_prob is being computed. Is deal_probability some CTR metric combined with some model Avito has on the relationship between CTR and deal_prob?
\n\nAlso, I noticed on the site that you must click a button to expose the sellers phone number. Is the number of clicks on this button how deal_prob is computed?
", "messages": [ { "id": 323593, "postDate": "2018-05-05T16:20:44.727Z", "content": "I still have not seen an adequate explanation of how deal_prob is being computed. Is deal_probability some CTR metric combined with some model Avito has on the relationship between CTR and deal_prob?
\n\nAlso, I noticed on the site that you must click a button to expose the sellers phone number. Is the number of clicks on this button how deal_prob is computed?
", "rawMarkdown": "I still have not seen an adequate explanation of how deal_prob is being computed. Is deal_probability some CTR metric combined with some model Avito has on the relationship between CTR and deal_prob? \n\nAlso, I noticed on the site that you must click a button to expose the sellers phone number. Is the number of clicks on this button how deal_prob is computed? ", "votes": 6 }, { "id": 324016, "postDate": "2018-05-06T23:45:48.523Z", "content": "I have exactly the same concern as you, and worry that this competition is not statistical enough. If deal_prob was created handcraftedly or it depends on some other factors rather than the raw features they give us, then who with the most powerful machine and the most time to test and combine huge model will be the winner by a very narrow gap. This would be unfair for some of us who focus on feature engineering.
", "rawMarkdown": "I have exactly the same concern as you, and worry that this competition is not statistical enough. If deal_prob was created handcraftedly or it depends on some other factors rather than the raw features they give us, then who with the most powerful machine and the most time to test and combine huge model will be the winner by a very narrow gap. This would be unfair for some of us who focus on feature engineering.", "votes": 1, "replies": [ { "id": 327382, "postDate": "2018-05-11T12:08:11.473Z", "content": "Generally it would be nice to have competitions where compute is restricted to kaggle kernel, i.e., you must submit from a kernel which would be helpful to level the playing field and reward insightful feature engineering and computational optimization.
", "rawMarkdown": "Generally it would be nice to have competitions where compute is restricted to kaggle kernel, i.e., you must submit from a kernel which would be helpful to level the playing field and reward insightful feature engineering and computational optimization. ", "votes": 2 }, { "id": 329993, "postDate": "2018-05-17T19:16:19.870Z", "content": "@ marketneutral - I agree. hated that approach when I worked on Mercari but I realized that its the best way to innovate, find new libraries and build new solutions as approached to just blending. Not opposed to blending but just saying kernel only helps all of us think out of the box a little more
", "rawMarkdown": "@ marketneutral - I agree. hated that approach when I worked on Mercari but I realized that its the best way to innovate, find new libraries and build new solutions as approached to just blending. Not opposed to blending but just saying kernel only helps all of us think out of the box a little more" } ] }, { "id": 327330, "postDate": "2018-05-11T09:52:36.450Z", "content": "Look at here https://www.kaggle.com/c/avito-demand-prediction/discussion/56531\nSome nice consideration.
\n\nBy the way, almost in all competitions your result depends on what machine you have
", "rawMarkdown": "Look at here https://www.kaggle.com/c/avito-demand-prediction/discussion/56531\nSome nice consideration.\n\nBy the way, almost in all competitions your result depends on what machine you have" }, { "id": 324002, "postDate": "2018-05-06T22:21:03.117Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 324016, "author_name": "Kha Vo", "author_url": "", "post_date": "2018-05-06T23:45:48.523000", "content": "I have exactly the same concern as you, and worry that this competition is not statistical enough. If deal_prob was created handcraftedly or it depends on some other factors rather than the raw features they give us, then who with the most powerful machine and the most time to test and combine huge model will be the winner by a very narrow gap. This would be unfair for some of us who focus on feature engineering.
", "votes": 1, "replies": [ { "id": 327382, "author_name": "marketneutral", "author_url": "", "post_date": "2018-05-11T12:08:11.473000", "content": "Generally it would be nice to have competitions where compute is restricted to kaggle kernel, i.e., you must submit from a kernel which would be helpful to level the playing field and reward insightful feature engineering and computational optimization.
", "votes": 2, "replies": [] }, { "id": 329993, "author_name": "Shanth", "author_url": "", "post_date": "2018-05-17T19:16:19.870000", "content": "@ marketneutral - I agree. hated that approach when I worked on Mercari but I realized that its the best way to innovate, find new libraries and build new solutions as approached to just blending. Not opposed to blending but just saying kernel only helps all of us think out of the box a little more
", "votes": 0, "replies": [] } ] }, { "id": 327330, "author_name": "Insaf Ashrapov", "author_url": "", "post_date": "2018-05-11T09:52:36.450000", "content": "Look at here https://www.kaggle.com/c/avito-demand-prediction/discussion/56531\nSome nice consideration.
\n\nBy the way, almost in all competitions your result depends on what machine you have
", "votes": 0, "replies": [] }, { "id": 324002, "author_name": "", "author_url": "", "post_date": "2018-05-06T22:21:03.117000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "323593": "I still have not seen an adequate explanation of how deal_prob is being computed. Is deal_probability some CTR metric combined with some model Avito has on the relationship between CTR and deal_prob? \n\nAlso, I noticed on the site that you must click a button to expose the sellers phone number. Is the number of clicks on this button how deal_prob is computed? ", "324016": "I have exactly the same concern as you, and worry that this competition is not statistical enough. If deal_prob was created handcraftedly or it depends on some other factors rather than the raw features they give us, then who with the most powerful machine and the most time to test and combine huge model will be the winner by a very narrow gap. This would be unfair for some of us who focus on feature engineering.", "327330": "Look at here https://www.kaggle.com/c/avito-demand-prediction/discussion/56531\nSome nice consideration.\n\nBy the way, almost in all competitions your result depends on what machine you have", "324002": "" } }