I am a noob in kaggle and discover it this year as I had a course on machine learning competition. I would like to express my gratitude and joy for having participated to this nice competition.
\n\n** Thanks to Kaggle and to Talking Data ** for organizing this nice competition\nI would like to share with you some insights:
\n\nI learned the hard way that you need to be careful about your resources and manage memory efficiently (see my post This competition is about RAM access. Before it, I never bothered to call gc.collet() or delete my variables, or set up the swap space on my computers... I managed to run the full dataset on my 16G computer, which at first I would never have imagined possible
I realized feature engeeneering is the core of the competition. I takes a lot of time to find the good features at least for me but it is definitely worth it.
It is very important to follow the competition in the last day. This is where things can change dramatically for newbies like me. It will not hurt top competitors but it is quite important for novices.
I discover lightgbm. I am a sklearn enthusiastic and I have spent a lot of time at the beginning using random forest. This was by far beaten by lightgbm.
I would like to thank the kaggle community. It is nice reading post of kagglers and exchange. I like the overall spirit of sharing
I will try to participate to other competitions if my wife and children permit\nAll the best to Kagglers and Kaggle!
", "messages": [ { "id": "325125", "postDate": "05/08/2018 06:34:14", "content": "I am a noob in kaggle and discover it this year as I had a course on machine learning competition. I would like to express my gratitude and joy for having participated to this nice competition.
\n\n** Thanks to Kaggle and to Talking Data ** for organizing this nice competition\nI would like to share with you some insights:
\n\nI learned the hard way that you need to be careful about your resources and manage memory efficiently (see my post This competition is about RAM access. Before it, I never bothered to call gc.collet() or delete my variables, or set up the swap space on my computers... I managed to run the full dataset on my 16G computer, which at first I would never have imagined possible
I realized feature engeeneering is the core of the competition. I takes a lot of time to find the good features at least for me but it is definitely worth it.
It is very important to follow the competition in the last day. This is where things can change dramatically for newbies like me. It will not hurt top competitors but it is quite important for novices.
I discover lightgbm. I am a sklearn enthusiastic and I have spent a lot of time at the beginning using random forest. This was by far beaten by lightgbm.
I would like to thank the kaggle community. It is nice reading post of kagglers and exchange. I like the overall spirit of sharing
I will try to participate to other competitions if my wife and children permit\nAll the best to Kagglers and Kaggle!
", "rawMarkdown": "I am a noob in kaggle and discover it this year as I had a course on machine learning competition. I would like to express my gratitude and joy for having participated to this nice competition.\n\n** Thanks to Kaggle and to Talking Data ** for organizing this nice competition\nI would like to share with you some insights:\n\n1. I learned the hard way that you need to be careful about your resources and manage memory efficiently (see my post [This competition is about RAM access](https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/56208). Before it, I never bothered to call gc.collet() or delete my variables, or set up the swap space on my computers... I managed to run the full dataset on my 16G computer, which at first I would never have imagined possible\n\n2. I realized feature engeeneering is the core of the competition. I takes a lot of time to find the good features at least for me but it is definitely worth it.\n\n3. It is very important to follow the competition in the last day. This is where things can change dramatically for newbies like me. It will not hurt top competitors but it is quite important for novices.\n\n4. I discover lightgbm. I am a sklearn enthusiastic and I have spent a lot of time at the beginning using random forest. This was by far beaten by lightgbm.\n\n5. I would like to thank the kaggle community. It is nice reading post of kagglers and exchange. I like the overall spirit of sharing\n\nI will try to participate to other competitions if my wife and children permit\nAll the best to Kagglers and Kaggle!", "votes": null }, { "id": "325772", "postDate": "05/08/2018 21:00:32", "content": "In addition to my comment, I would like to express my gratitude to top kagglers fro sharing some of their tricks.
\n\nSpecial thanks to
\n\nbestfitting who shares insights\nabout his solution with 3rd position and some hints about Neural networks : My brief summary, a\nmainly NN based solution
CPMP : who as usual provided lots of details about his solution and in particular interesting remarks on matrix factorization Solution #6 overview
FaiYang with some hinsights about doing a dot product in a way similar to libfm..
I did not have the time to read all now but will spend more time reading from the experts :-)
", "rawMarkdown": "In addition to my comment, I would like to express my gratitude to top kagglers fro sharing some of their tricks. \n\n**Special thanks to**\n\n * [bestfitting](https://www.kaggle.com/bestfitting) who shares insights\n about his solution with 3rd position and some hints about Neural networks : [My brief summary, a\n mainly NN based solution](https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/56262#latest-325766)\n\n * [CPMP](https://www.kaggle.com/cpmpml) : who as usual provided lots of details about his solution and in particular interesting remarks on matrix factorization [Solution #6 overview](https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/56283#latest-325749)\n\n * [Kazanova](https://www.kaggle.com/kazanova) [4th place (brief) tips](https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/56243#latest-325716)\n\n * [FaiYang](https://www.kaggle.com/panfeiyang) with some hinsights about doing a dot product in a way similar to libfm..\n \nI did not have the time to read all now but will spend more time reading from the **experts** :-)", "votes": null }, { "id": "325790", "postDate": "05/08/2018 21:25:30", "content": "Yes, this was a great competition! So many wonderful, standout contributors here on Kaggle who share their knowledge so openly, and you've named many of them. The only downer was that numerous people just submitted Dirk's results as-is and \"won\" medals. That takes the fun out of competing. When one is doing data science in the real world, there's no CSV you can download and submit :-) Haha.
", "rawMarkdown": "Yes, this was a great competition! So many wonderful, standout contributors here on Kaggle who share their knowledge so openly, and you've named many of them. The only downer was that numerous people just submitted Dirk's results as-is and \"won\" medals. That takes the fun out of competing. When one is doing data science in the real world, there's no CSV you can download and submit :-) Haha.", "votes": null } ], "comments": [ { "id": 325772, "author_name": "ericbenhamou", "author_url": "", "post_date": "05/08/2018 21:00:32", "content": "In addition to my comment, I would like to express my gratitude to top kagglers fro sharing some of their tricks.
\n\nSpecial thanks to
\n\nbestfitting who shares insights\nabout his solution with 3rd position and some hints about Neural networks : My brief summary, a\nmainly NN based solution
CPMP : who as usual provided lots of details about his solution and in particular interesting remarks on matrix factorization Solution #6 overview
FaiYang with some hinsights about doing a dot product in a way similar to libfm..
I did not have the time to read all now but will spend more time reading from the experts :-)
", "votes": null, "replies": [] }, { "id": 325790, "author_name": "eigenvector", "author_url": "", "post_date": "05/08/2018 21:25:30", "content": "Yes, this was a great competition! So many wonderful, standout contributors here on Kaggle who share their knowledge so openly, and you've named many of them. The only downer was that numerous people just submitted Dirk's results as-is and \"won\" medals. That takes the fun out of competing. When one is doing data science in the real world, there's no CSV you can download and submit :-) Haha.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "325125": "I am a noob in kaggle and discover it this year as I had a course on machine learning competition. I would like to express my gratitude and joy for having participated to this nice competition.\n\n** Thanks to Kaggle and to Talking Data ** for organizing this nice competition\nI would like to share with you some insights:\n\n1. I learned the hard way that you need to be careful about your resources and manage memory efficiently (see my post [This competition is about RAM access](https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/56208). Before it, I never bothered to call gc.collet() or delete my variables, or set up the swap space on my computers... I managed to run the full dataset on my 16G computer, which at first I would never have imagined possible\n\n2. I realized feature engeeneering is the core of the competition. I takes a lot of time to find the good features at least for me but it is definitely worth it.\n\n3. It is very important to follow the competition in the last day. This is where things can change dramatically for newbies like me. It will not hurt top competitors but it is quite important for novices.\n\n4. I discover lightgbm. I am a sklearn enthusiastic and I have spent a lot of time at the beginning using random forest. This was by far beaten by lightgbm.\n\n5. I would like to thank the kaggle community. It is nice reading post of kagglers and exchange. I like the overall spirit of sharing\n\nI will try to participate to other competitions if my wife and children permit\nAll the best to Kagglers and Kaggle!", "325772": "In addition to my comment, I would like to express my gratitude to top kagglers fro sharing some of their tricks. \n\n**Special thanks to**\n\n * [bestfitting](https://www.kaggle.com/bestfitting) who shares insights\n about his solution with 3rd position and some hints about Neural networks : [My brief summary, a\n mainly NN based solution](https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/56262#latest-325766)\n\n * [CPMP](https://www.kaggle.com/cpmpml) : who as usual provided lots of details about his solution and in particular interesting remarks on matrix factorization [Solution #6 overview](https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/56283#latest-325749)\n\n * [Kazanova](https://www.kaggle.com/kazanova) [4th place (brief) tips](https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/56243#latest-325716)\n\n * [FaiYang](https://www.kaggle.com/panfeiyang) with some hinsights about doing a dot product in a way similar to libfm..\n \nI did not have the time to read all now but will spend more time reading from the **experts** :-)", "325790": "Yes, this was a great competition! So many wonderful, standout contributors here on Kaggle who share their knowledge so openly, and you've named many of them. The only downer was that numerous people just submitted Dirk's results as-is and \"won\" medals. That takes the fun out of competing. When one is doing data science in the real world, there's no CSV you can download and submit :-) Haha." }, "source": "meta" }