Hi all\nI am new here to learn about machine learning on real-world problems. Probably like many others too. Sorry, if this is a bit off-topic or has been discussed already. My question is a rather general one.\nI understand that this is a real-world problem with real data from Elo.\nAnd my understanding so far is that data understanding and preparation is usually the most important and most time consuming part for such problems: data exploration, cleaning, feature engineering.\nAfter that it's actually a matter of selecting an appropriate learning algorithm (e.g. a sophisticated decision tree like lightgbm) and let it train on the features engineered before hand.\nPlease correct me, if I am fundamentally wrong here.
\n\nNow to my question: \nSo this is a real-world problem but I am wondering how \"real-world\" is the competition?\nIn your experience on off-competition machine learning problems, does it really matter if the RMSE is 3.9 or 3.6? Or even smaller differences.
\n\nBroader Question:\nIs there an important learning effect in tweaking the features and parameters of the learning algorithm? Is it worth the time or should I rather invest my time in exploring and understanding several different kind of problems?
\n\nMany thanks for your answers!
", "messages": [ { "id": "461088", "postDate": "01/25/2019 07:35:19", "content": "Hi all\nI am new here to learn about machine learning on real-world problems. Probably like many others too. Sorry, if this is a bit off-topic or has been discussed already. My question is a rather general one.\nI understand that this is a real-world problem with real data from Elo.\nAnd my understanding so far is that data understanding and preparation is usually the most important and most time consuming part for such problems: data exploration, cleaning, feature engineering.\nAfter that it's actually a matter of selecting an appropriate learning algorithm (e.g. a sophisticated decision tree like lightgbm) and let it train on the features engineered before hand.\nPlease correct me, if I am fundamentally wrong here.
\n\nNow to my question: \nSo this is a real-world problem but I am wondering how \"real-world\" is the competition?\nIn your experience on off-competition machine learning problems, does it really matter if the RMSE is 3.9 or 3.6? Or even smaller differences.
\n\nBroader Question:\nIs there an important learning effect in tweaking the features and parameters of the learning algorithm? Is it worth the time or should I rather invest my time in exploring and understanding several different kind of problems?
\n\nMany thanks for your answers!
", "rawMarkdown": "Hi all\nI am new here to learn about machine learning on real-world problems. Probably like many others too. Sorry, if this is a bit off-topic or has been discussed already. My question is a rather general one.\nI understand that this is a real-world problem with real data from Elo.\nAnd my understanding so far is that data understanding and preparation is usually the most important and most time consuming part for such problems: data exploration, cleaning, feature engineering.\nAfter that it's actually a matter of selecting an appropriate learning algorithm (e.g. a sophisticated decision tree like lightgbm) and let it train on the features engineered before hand.\nPlease correct me, if I am fundamentally wrong here.\n\nNow to my question: \nSo this is a real-world problem but I am wondering how \"real-world\" is the competition?\nIn your experience on off-competition machine learning problems, does it really matter if the RMSE is 3.9 or 3.6? Or even smaller differences.\n\nBroader Question:\nIs there an important learning effect in tweaking the features and parameters of the learning algorithm? Is it worth the time or should I rather invest my time in exploring and understanding several different kind of problems?\n\nMany thanks for your answers!", "votes": null }, { "id": "462227", "postDate": "01/27/2019 22:54:16", "content": "I had asked the same questions to a \"real world\" Data Scientist and his response was something like this:\n1) Data Scientists spend most of their time preparing the Data set which is different from Kaggle competitions.\n2) They do have a threshold for RMSE and other scores but do not spend too much time in parameter tuning once they have reached the threshold or close to it. (we are talking about ML in business, not in health)\n3) Companies rarely look at your Kaggle profile unless you are ranked in the top 10/20.
\n\nI had asked a friend whose team mainly consist of Data Scientists and she said that they do not look at Kaggle profile at all.
\n\nI personally, however, have benefitted a lot by following Kaggle competitions and the discussion board. When I started working on it a couple of years ago, these helped me design real-world ML project at work, build data set with real-world data for various scenarios. It really helps newbies build their knowledge base and confidence in ML domain.
", "rawMarkdown": "I had asked the same questions to a \"real world\" Data Scientist and his response was something like this:\n1) Data Scientists spend most of their time preparing the Data set which is different from Kaggle competitions.\n2) They do have a threshold for RMSE and other scores but do not spend too much time in parameter tuning once they have reached the threshold or close to it. (we are talking about ML in business, not in health)\n3) Companies rarely look at your Kaggle profile unless you are ranked in the top 10/20.\n\nI had asked a friend whose team mainly consist of Data Scientists and she said that they do not look at Kaggle profile at all.\n\nI personally, however, have benefitted a lot by following Kaggle competitions and the discussion board. When I started working on it a couple of years ago, these helped me design real-world ML project at work, build data set with real-world data for various scenarios. It really helps newbies build their knowledge base and confidence in ML domain.", "votes": null }, { "id": "462598", "postDate": "01/28/2019 15:11:13", "content": "Thank you very much for you reply. That helps a lot!
", "rawMarkdown": "Thank you very much for you reply. That helps a lot!", "votes": null } ], "comments": [ { "id": 462227, "author_name": "kanjasaha", "author_url": "", "post_date": "01/27/2019 22:54:16", "content": "I had asked the same questions to a \"real world\" Data Scientist and his response was something like this:\n1) Data Scientists spend most of their time preparing the Data set which is different from Kaggle competitions.\n2) They do have a threshold for RMSE and other scores but do not spend too much time in parameter tuning once they have reached the threshold or close to it. (we are talking about ML in business, not in health)\n3) Companies rarely look at your Kaggle profile unless you are ranked in the top 10/20.
\n\nI had asked a friend whose team mainly consist of Data Scientists and she said that they do not look at Kaggle profile at all.
\n\nI personally, however, have benefitted a lot by following Kaggle competitions and the discussion board. When I started working on it a couple of years ago, these helped me design real-world ML project at work, build data set with real-world data for various scenarios. It really helps newbies build their knowledge base and confidence in ML domain.
", "votes": null, "replies": [ { "id": 462598, "author_name": "jaminalder", "author_url": "", "post_date": "01/28/2019 15:11:13", "content": "Thank you very much for you reply. That helps a lot!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "461088": "Hi all\nI am new here to learn about machine learning on real-world problems. Probably like many others too. Sorry, if this is a bit off-topic or has been discussed already. My question is a rather general one.\nI understand that this is a real-world problem with real data from Elo.\nAnd my understanding so far is that data understanding and preparation is usually the most important and most time consuming part for such problems: data exploration, cleaning, feature engineering.\nAfter that it's actually a matter of selecting an appropriate learning algorithm (e.g. a sophisticated decision tree like lightgbm) and let it train on the features engineered before hand.\nPlease correct me, if I am fundamentally wrong here.\n\nNow to my question: \nSo this is a real-world problem but I am wondering how \"real-world\" is the competition?\nIn your experience on off-competition machine learning problems, does it really matter if the RMSE is 3.9 or 3.6? Or even smaller differences.\n\nBroader Question:\nIs there an important learning effect in tweaking the features and parameters of the learning algorithm? Is it worth the time or should I rather invest my time in exploring and understanding several different kind of problems?\n\nMany thanks for your answers!", "462227": "I had asked the same questions to a \"real world\" Data Scientist and his response was something like this:\n1) Data Scientists spend most of their time preparing the Data set which is different from Kaggle competitions.\n2) They do have a threshold for RMSE and other scores but do not spend too much time in parameter tuning once they have reached the threshold or close to it. (we are talking about ML in business, not in health)\n3) Companies rarely look at your Kaggle profile unless you are ranked in the top 10/20.\n\nI had asked a friend whose team mainly consist of Data Scientists and she said that they do not look at Kaggle profile at all.\n\nI personally, however, have benefitted a lot by following Kaggle competitions and the discussion board. When I started working on it a couple of years ago, these helped me design real-world ML project at work, build data set with real-world data for various scenarios. It really helps newbies build their knowledge base and confidence in ML domain.", "462598": "Thank you very much for you reply. That helps a lot!" }, "source": "meta" }