I have read a lot of kernels dealing about feature engineering and EDA. However I wasn't able to find any using Mulliken charges and potential energy as features for a model. I'm curious if you've tried something similar.
\n\nRight now I'm planning on including Mulliken charges and potential energy in my model as simple features...
", "messages": [ { "id": "562726", "postDate": "06/27/2019 13:49:13", "content": "I have read a lot of kernels dealing about feature engineering and EDA. However I wasn't able to find any using Mulliken charges and potential energy as features for a model. I'm curious if you've tried something similar.
\n\nRight now I'm planning on including Mulliken charges and potential energy in my model as simple features...
", "rawMarkdown": "I have read a lot of kernels dealing about feature engineering and EDA. However I wasn't able to find any using Mulliken charges and potential energy as features for a model. I'm curious if you've tried something similar. \n\nRight now I'm planning on including Mulliken charges and potential energy in my model as simple features...", "votes": null }, { "id": "562753", "postDate": "06/27/2019 14:09:20", "content": "The external datasets / features like mulliken charges and potential energy are only provided for training data. So, if you use it directly by merging, you are bound to get good local CV scores but those features will be missing for test data.\nYou can look at Andrew's kernel for ideas on how they can be used appropriately as meta features for your model so that it is available for test data as well without any leakage/missing values.
", "rawMarkdown": "The external datasets / features like mulliken charges and potential energy are only provided for training data. So, if you use it directly by merging, you are bound to get good local CV scores but those features will be missing for test data.\nYou can look at Andrew's [kernel](https://www.kaggle.com/artgor/using-meta-features-to-improve-model) for ideas on how they can be used appropriately as meta features for your model so that it is available for test data as well without any leakage/missing values.", "votes": null }, { "id": "562846", "postDate": "06/27/2019 15:35:26", "content": "Two kernels that used the Mulliken charges for training: here and here and the comparison of both approaches: here.
", "rawMarkdown": "Two kernels that used the Mulliken charges for training: [here](https://www.kaggle.com/borisdee/predicting-mulliken-charges-with-acsf-descriptors) and [here](https://www.kaggle.com/asauve/v7-estimation-of-mulliken-charges-with-open-babel) and the comparison of both approaches: [here](https://www.kaggle.com/asauve/comparison-of-acsf-built-charges-with-open-babel).", "votes": null }, { "id": "563021", "postDate": "06/27/2019 19:08:29", "content": "Another kernel using Mulliken charges and potential energy.
", "rawMarkdown": "Another [kernel using Mulliken charges and potential energy](https://www.kaggle.com/robertburbidge/using-estimated-mulliken-charges).", "votes": null }, { "id": "563577", "postDate": "06/28/2019 12:58:37", "content": "@rohanrao I will check it! Thanks!
", "rawMarkdown": "rohanrao I will check it! Thanks!", "votes": null } ], "comments": [ { "id": 562753, "author_name": "rohanrao", "author_url": "", "post_date": "06/27/2019 14:09:20", "content": "The external datasets / features like mulliken charges and potential energy are only provided for training data. So, if you use it directly by merging, you are bound to get good local CV scores but those features will be missing for test data.\nYou can look at Andrew's kernel for ideas on how they can be used appropriately as meta features for your model so that it is available for test data as well without any leakage/missing values.
", "votes": null, "replies": [] }, { "id": 562846, "author_name": "borisdee", "author_url": "", "post_date": "06/27/2019 15:35:26", "content": "Two kernels that used the Mulliken charges for training: here and here and the comparison of both approaches: here.
", "votes": null, "replies": [] }, { "id": 563021, "author_name": "robertburbidge", "author_url": "", "post_date": "06/27/2019 19:08:29", "content": "Another kernel using Mulliken charges and potential energy.
", "votes": null, "replies": [] }, { "id": 563577, "author_name": "rftexas", "author_url": "", "post_date": "06/28/2019 12:58:37", "content": "@rohanrao I will check it! Thanks!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "562726": "I have read a lot of kernels dealing about feature engineering and EDA. However I wasn't able to find any using Mulliken charges and potential energy as features for a model. I'm curious if you've tried something similar. \n\nRight now I'm planning on including Mulliken charges and potential energy in my model as simple features...", "562753": "The external datasets / features like mulliken charges and potential energy are only provided for training data. So, if you use it directly by merging, you are bound to get good local CV scores but those features will be missing for test data.\nYou can look at Andrew's [kernel](https://www.kaggle.com/artgor/using-meta-features-to-improve-model) for ideas on how they can be used appropriately as meta features for your model so that it is available for test data as well without any leakage/missing values.", "562846": "Two kernels that used the Mulliken charges for training: [here](https://www.kaggle.com/borisdee/predicting-mulliken-charges-with-acsf-descriptors) and [here](https://www.kaggle.com/asauve/v7-estimation-of-mulliken-charges-with-open-babel) and the comparison of both approaches: [here](https://www.kaggle.com/asauve/comparison-of-acsf-built-charges-with-open-babel).", "563021": "Another [kernel using Mulliken charges and potential energy](https://www.kaggle.com/robertburbidge/using-estimated-mulliken-charges).", "563577": "rohanrao I will check it! Thanks!" }, "source": "meta" }