basic question --> almost all the posted kernels are using the tuned params from Bayesian optimization. But do we tune params again on every-time new feature/set of features are introduced or cut down? based on past experience can you share what are the ideal scenarios to retune the params?
", "messages": [ { "id": "459042", "postDate": "01/21/2019 04:18:33", "content": "basic question --> almost all the posted kernels are using the tuned params from Bayesian optimization. But do we tune params again on every-time new feature/set of features are introduced or cut down? based on past experience can you share what are the ideal scenarios to retune the params?
", "rawMarkdown": "basic question --> almost all the posted kernels are using the tuned params from Bayesian optimization. But do we tune params again on every-time new feature/set of features are introduced or cut down? based on past experience can you share what are the ideal scenarios to retune the params?", "votes": null }, { "id": "459101", "postDate": "01/21/2019 07:22:31", "content": "I think that the parameter tuning should be the final stage of model development. First you need to understand the data and do a cleanup and preprocessing. For the feature-engineering is better to stick with one model parameter settings. Then you can compare the CV and LB score with the previous data version and evaluate the impact of a new feature. And when your data are ready, then you can start tuning the parameters. Tuning parameters after every change will be time-consuming especially in the case of this large dataset.
", "rawMarkdown": "I think that the parameter tuning should be the final stage of model development. First you need to understand the data and do a cleanup and preprocessing. For the feature-engineering is better to stick with one model parameter settings. Then you can compare the CV and LB score with the previous data version and evaluate the impact of a new feature. And when your data are ready, then you can start tuning the parameters. Tuning parameters after every change will be time-consuming especially in the case of this large dataset.", "votes": null }, { "id": "459336", "postDate": "01/21/2019 15:42:05", "content": "Thanks for the reply! this makes sense - set the common ground for different versions of datasets, models and then pick the best one to refine further in terms of parameter tuning.
", "rawMarkdown": "Thanks for the reply! this makes sense - set the common ground for different versions of datasets, models and then pick the best one to refine further in terms of parameter tuning.", "votes": null }, { "id": "459697", "postDate": "01/22/2019 07:34:58", "content": "Yes, train several models is a good strategy. If you have several good models you can then create a simple blend of their results. A simple average of their predicted probabilities is a good start. This usually leads to better score, especially if the models are not correlated.
", "rawMarkdown": "Yes, train several models is a good strategy. If you have several good models you can then create a simple blend of their results. A simple average of their predicted probabilities is a good start. This usually leads to better score, especially if the models are not correlated.", "votes": null }, { "id": "459886", "postDate": "01/22/2019 15:14:14", "content": "Hi everybody, i'm new in the competition and i'm wondering how to deal with the transactions database.\nMy computer hasn't enough memory for this database.
", "rawMarkdown": "Hi everybody, i'm new in the competition and i'm wondering how to deal with the transactions database.\nMy computer hasn't enough memory for this database.", "votes": null }, { "id": "459940", "postDate": "01/22/2019 16:52:59", "content": "Hi Toto, please search through the discussion threads for the relevant thread or start a new one if you don’t find anything pertaining to your query. Let me redirect you to thread similar to your question- https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/75713
", "rawMarkdown": "Hi Toto, please search through the discussion threads for the relevant thread or start a new one if you don’t find anything pertaining to your query. Let me redirect you to thread similar to your question- https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/75713", "votes": null } ], "comments": [ { "id": 459101, "author_name": "jirivrany", "author_url": "", "post_date": "01/21/2019 07:22:31", "content": "I think that the parameter tuning should be the final stage of model development. First you need to understand the data and do a cleanup and preprocessing. For the feature-engineering is better to stick with one model parameter settings. Then you can compare the CV and LB score with the previous data version and evaluate the impact of a new feature. And when your data are ready, then you can start tuning the parameters. Tuning parameters after every change will be time-consuming especially in the case of this large dataset.
", "votes": null, "replies": [ { "id": 459336, "author_name": "apoorvg", "author_url": "", "post_date": "01/21/2019 15:42:05", "content": "Thanks for the reply! this makes sense - set the common ground for different versions of datasets, models and then pick the best one to refine further in terms of parameter tuning.
", "votes": null, "replies": [] }, { "id": 459697, "author_name": "jirivrany", "author_url": "", "post_date": "01/22/2019 07:34:58", "content": "Yes, train several models is a good strategy. If you have several good models you can then create a simple blend of their results. A simple average of their predicted probabilities is a good start. This usually leads to better score, especially if the models are not correlated.
", "votes": null, "replies": [] } ] }, { "id": 459886, "author_name": "elimanealain", "author_url": "", "post_date": "01/22/2019 15:14:14", "content": "Hi everybody, i'm new in the competition and i'm wondering how to deal with the transactions database.\nMy computer hasn't enough memory for this database.
", "votes": null, "replies": [ { "id": 459940, "author_name": "apoorvg", "author_url": "", "post_date": "01/22/2019 16:52:59", "content": "Hi Toto, please search through the discussion threads for the relevant thread or start a new one if you don’t find anything pertaining to your query. Let me redirect you to thread similar to your question- https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/75713
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "459042": "basic question --> almost all the posted kernels are using the tuned params from Bayesian optimization. But do we tune params again on every-time new feature/set of features are introduced or cut down? based on past experience can you share what are the ideal scenarios to retune the params?", "459101": "I think that the parameter tuning should be the final stage of model development. First you need to understand the data and do a cleanup and preprocessing. For the feature-engineering is better to stick with one model parameter settings. Then you can compare the CV and LB score with the previous data version and evaluate the impact of a new feature. And when your data are ready, then you can start tuning the parameters. Tuning parameters after every change will be time-consuming especially in the case of this large dataset.", "459336": "Thanks for the reply! this makes sense - set the common ground for different versions of datasets, models and then pick the best one to refine further in terms of parameter tuning.", "459697": "Yes, train several models is a good strategy. If you have several good models you can then create a simple blend of their results. A simple average of their predicted probabilities is a good start. This usually leads to better score, especially if the models are not correlated.", "459886": "Hi everybody, i'm new in the competition and i'm wondering how to deal with the transactions database.\nMy computer hasn't enough memory for this database.", "459940": "Hi Toto, please search through the discussion threads for the relevant thread or start a new one if you don’t find anything pertaining to your query. Let me redirect you to thread similar to your question- https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/75713" }, "source": "meta" }