I'm a little stuck in my efforts to get a better prediction, and I figure I need to include merchants.csv at this stage. Anyone tried deep feature synthesis yet?
", "messages": [ { "id": "431261", "postDate": "12/02/2018 00:41:29", "content": "I'm a little stuck in my efforts to get a better prediction, and I figure I need to include merchants.csv at this stage. Anyone tried deep feature synthesis yet?
", "rawMarkdown": "I'm a little stuck in my efforts to get a better prediction, and I figure I need to include merchants.csv at this stage. Anyone tried deep feature synthesis yet?", "votes": null }, { "id": "431782", "postDate": "12/02/2018 22:14:18", "content": "No DL so far, but I did try using merchant-based features. Improved CV, decreased LB score - which is pretty much a recurring theme for me in this competition so far.
", "rawMarkdown": "No DL so far, but I did try using merchant-based features. Improved CV, decreased LB score - which is pretty much a recurring theme for me in this competition so far.", "votes": null }, { "id": "431786", "postDate": "12/02/2018 22:23:31", "content": "I've been struggling to get anywhere with hyperparameter tunings of LightGBM. Feature engg is my only way out it seems.
", "rawMarkdown": "I've been struggling to get anywhere with hyperparameter tunings of LightGBM. Feature engg is my only way out it seems.", "votes": null }, { "id": "433738", "postDate": "12/05/2018 12:34:16", "content": "Feature engineering is the aspect that will give you biggest \"jumps\" in score from my experience (in machine learning). Hyperparameter is not where you will get huge differences. If you also look at top 10 kernels on other competitions where Machine Learning was used, you'll see that FE and Stacking are usually the solutions
", "rawMarkdown": "Feature engineering is the aspect that will give you biggest \"jumps\" in score from my experience (in machine learning). Hyperparameter is not where you will get huge differences. If you also look at top 10 kernels on other competitions where Machine Learning was used, you'll see that FE and Stacking are usually the solutions", "votes": null }, { "id": "433988", "postDate": "12/05/2018 18:29:31", "content": "You may simply join merchants.csv with new_merchant_transactions.csv and create some aggregations.
", "rawMarkdown": "You may simply join merchants.csv with new\\_merchant\\_transactions.csv and create some aggregations.", "votes": null }, { "id": "433998", "postDate": "12/05/2018 18:46:22", "content": "I already tried.. https://www.kaggle.com/delayedkarma/utilizing-features-from-merchants-csv.. to no avail
", "rawMarkdown": "I already tried.. https://www.kaggle.com/delayedkarma/utilizing-features-from-merchants-csv.. to no avail", "votes": null } ], "comments": [ { "id": 431782, "author_name": "konradb", "author_url": "", "post_date": "12/02/2018 22:14:18", "content": "No DL so far, but I did try using merchant-based features. Improved CV, decreased LB score - which is pretty much a recurring theme for me in this competition so far.
", "votes": null, "replies": [ { "id": 431786, "author_name": "delayedkarma", "author_url": "", "post_date": "12/02/2018 22:23:31", "content": "I've been struggling to get anywhere with hyperparameter tunings of LightGBM. Feature engg is my only way out it seems.
", "votes": null, "replies": [] }, { "id": 433738, "author_name": "xtelima", "author_url": "", "post_date": "12/05/2018 12:34:16", "content": "Feature engineering is the aspect that will give you biggest \"jumps\" in score from my experience (in machine learning). Hyperparameter is not where you will get huge differences. If you also look at top 10 kernels on other competitions where Machine Learning was used, you'll see that FE and Stacking are usually the solutions
", "votes": null, "replies": [] } ] }, { "id": 433988, "author_name": "kailex", "author_url": "", "post_date": "12/05/2018 18:29:31", "content": "You may simply join merchants.csv with new_merchant_transactions.csv and create some aggregations.
", "votes": null, "replies": [ { "id": 433998, "author_name": "delayedkarma", "author_url": "", "post_date": "12/05/2018 18:46:22", "content": "I already tried.. https://www.kaggle.com/delayedkarma/utilizing-features-from-merchants-csv.. to no avail
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "431261": "I'm a little stuck in my efforts to get a better prediction, and I figure I need to include merchants.csv at this stage. Anyone tried deep feature synthesis yet?", "431782": "No DL so far, but I did try using merchant-based features. Improved CV, decreased LB score - which is pretty much a recurring theme for me in this competition so far.", "431786": "I've been struggling to get anywhere with hyperparameter tunings of LightGBM. Feature engg is my only way out it seems.", "433738": "Feature engineering is the aspect that will give you biggest \"jumps\" in score from my experience (in machine learning). Hyperparameter is not where you will get huge differences. If you also look at top 10 kernels on other competitions where Machine Learning was used, you'll see that FE and Stacking are usually the solutions", "433988": "You may simply join merchants.csv with new\\_merchant\\_transactions.csv and create some aggregations.", "433998": "I already tried.. https://www.kaggle.com/delayedkarma/utilizing-features-from-merchants-csv.. to no avail" }, "source": "meta" }