Lags can often be powerful features in time based problems. As seen in discussions and notebooks, many people are using lagged features or lagged targets (myself included).
\nAs we are likely to have a gap between train and test this could potentially cause problems for people using lags.
\nI'd be curious to see whether people are still planning to use lags in their submissions?
", "messages": [ { "id": "1393605", "postDate": "07/19/2021 18:52:51", "content": "Lags can often be powerful features in time based problems. As seen in discussions and notebooks, many people are using lagged features or lagged targets (myself included).
\nAs we are likely to have a gap between train and test this could potentially cause problems for people using lags.
\nI'd be curious to see whether people are still planning to use lags in their submissions?
", "rawMarkdown": "Lags can often be powerful features in time based problems. As seen in discussions and notebooks, many people are using lagged features or lagged targets (myself included). \n\nAs we are likely to have a gap between train and test this could potentially cause problems for people using lags.\n\nI'd be curious to see whether people are still planning to use lags in their submissions?", "votes": null }, { "id": "1393681", "postDate": "07/19/2021 19:45:39", "content": "I'm planning on hedging by having one submission that relies on them and one that doesn't.
\nI'm trying to give the Kaggle staff as much of a benefit of the doubt as possible because I'm sure that it's really hard to run these competitions, but it seems just a little bizarre that they can't tell us if there will be continuity or not. As you point out, it does seem like they're really winking and nodding that there will be a gap, though.
\nOne positive to them updating in the manner they are is that it pushes us to do feature engineering in kernels as opposed to simply uploading a pretrained model and being done with it. I've always appreciated adding an extra twist like that, almost remotely reminiscent of the 2019 Big Data Bowl.
\nEDIT: @juliaelliott addressed this in a different thread. I think her point is fair about the difficulty of running a competition like this.
", "rawMarkdown": "I'm planning on hedging by having one submission that relies on them and one that doesn't. \n\nI'm trying to give the Kaggle staff as much of a benefit of the doubt as possible because I'm sure that it's really hard to run these competitions, but it seems just a little bizarre that they can't tell us if there will be continuity or not. As you point out, it does seem like they're really winking and nodding that there will be a gap, though.\n\nOne positive to them updating in the manner they are is that it pushes us to do feature engineering in kernels as opposed to simply uploading a pretrained model and being done with it. I've always appreciated adding an extra twist like that, almost remotely reminiscent of the 2019 Big Data Bowl.\n\nEDIT: @juliaelliott addressed this in a different thread. I think her point is fair about the difficulty of running a competition like this.", "votes": null } ], "comments": [ { "id": 1393681, "author_name": "marktenenholtz", "author_url": "", "post_date": "07/19/2021 19:45:39", "content": "I'm planning on hedging by having one submission that relies on them and one that doesn't.
\nI'm trying to give the Kaggle staff as much of a benefit of the doubt as possible because I'm sure that it's really hard to run these competitions, but it seems just a little bizarre that they can't tell us if there will be continuity or not. As you point out, it does seem like they're really winking and nodding that there will be a gap, though.
\nOne positive to them updating in the manner they are is that it pushes us to do feature engineering in kernels as opposed to simply uploading a pretrained model and being done with it. I've always appreciated adding an extra twist like that, almost remotely reminiscent of the 2019 Big Data Bowl.
\nEDIT: @juliaelliott addressed this in a different thread. I think her point is fair about the difficulty of running a competition like this.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1393605": "Lags can often be powerful features in time based problems. As seen in discussions and notebooks, many people are using lagged features or lagged targets (myself included). \n\nAs we are likely to have a gap between train and test this could potentially cause problems for people using lags.\n\nI'd be curious to see whether people are still planning to use lags in their submissions?", "1393681": "I'm planning on hedging by having one submission that relies on them and one that doesn't. \n\nI'm trying to give the Kaggle staff as much of a benefit of the doubt as possible because I'm sure that it's really hard to run these competitions, but it seems just a little bizarre that they can't tell us if there will be continuity or not. As you point out, it does seem like they're really winking and nodding that there will be a gap, though.\n\nOne positive to them updating in the manner they are is that it pushes us to do feature engineering in kernels as opposed to simply uploading a pretrained model and being done with it. I've always appreciated adding an extra twist like that, almost remotely reminiscent of the 2019 Big Data Bowl.\n\nEDIT: @juliaelliott addressed this in a different thread. I think her point is fair about the difficulty of running a competition like this." }, "source": "meta" }