We could train a SOM on all the training data, including the data we have only for the train set. After that we could make a simplified SOM with the data that we have for both the training and test set, and train it only on train set. Then we could give that SOM test data and see where it places the train data, and with that info we could see where the test data would go in the SOM of all features. We could then train a regression network using the coordinates of train data in the SOM of all features (including other data), and use that info for predicting the constant of the test set. Would this be worth the effort?
\n\nWhat I'm hoping to achieve with this is to use all the data we have for training set, and since SOMs would cluster training data with test data, that would give a prediction as to what other features for test set are, that we don't have.
", "messages": [ { "id": 540941, "postDate": "2019-06-01T12:04:45.367Z", "content": "We could train a SOM on all the training data, including the data we have only for the train set. After that we could make a simplified SOM with the data that we have for both the training and test set, and train it only on train set. Then we could give that SOM test data and see where it places the train data, and with that info we could see where the test data would go in the SOM of all features. We could then train a regression network using the coordinates of train data in the SOM of all features (including other data), and use that info for predicting the constant of the test set. Would this be worth the effort?
\n\nWhat I'm hoping to achieve with this is to use all the data we have for training set, and since SOMs would cluster training data with test data, that would give a prediction as to what other features for test set are, that we don't have.
", "rawMarkdown": "We could train a SOM on all the training data, including the data we have **only** for the train set. After that we could make a simplified SOM with the data that we have for both the training and test set, and train it only on train set. Then we could give that SOM test data and see where it places the train data, and with that info we could see where the test data would go in the SOM of all features. We could then train a regression network using the coordinates of train data in the SOM of all features (including other data), and use that info for predicting the constant of the test set. Would this be worth the effort?\n\nWhat I'm hoping to achieve with this is to use all the data we have for training set, and since SOMs would cluster training data with test data, that would give a prediction as to what other features for test set are, that we don't have.", "votes": -2 }, { "id": 550853, "postDate": "2019-06-12T04:54:45.943Z", "content": "Not sure why people downvote this (-4), its a personal hypothesis,\nmaybe true or false. It may not make sense to the reader or expert,\nit maybe crazy or the ideas or assumptions are completely wrong,\nhowever the only task is to prove or disprove it.
\n\nWe have a large blinded validation set, nobody knows \nthe answer (yet) so @mkozic just needs to prove it or disprove it.\nNo need to downvote unless some additional advice is given.
\n\nSOMs (Kohonen neural networks) were the holy grail of algorithms that\ncould solve everything in the 90s. Reading old papers and books\nthey literally could be used for everything (world peace, no hunger & healing cancer)
\n\nSee Google Scholar links for SOMs: \nhttps://scholar.google.be/scholar?hl=nl-be&q=som+Kohonen+neural+network+chemistry
\n\nSo @mkozic yes please prove or disprove your hypothesis.
", "rawMarkdown": "Not sure why people downvote this (-4), its a personal hypothesis,\nmaybe true or false. It may not make sense to the reader or expert,\nit maybe crazy or the ideas or assumptions are completely wrong,\nhowever the only task is to prove or disprove it.\n\nWe have a large blinded validation set, nobody knows \nthe answer (yet) so @mkozic just needs to prove it or disprove it.\nNo need to downvote unless some additional advice is given.\n\nSOMs (Kohonen neural networks) were the holy grail of algorithms that\ncould solve everything in the 90s. Reading old papers and books\nthey literally could be used for everything (world peace, no hunger & healing cancer)\n\nSee Google Scholar links for SOMs: \nhttps://scholar.google.be/scholar?hl=nl-be&q=som+Kohonen+neural+network+chemistry\n\nSo @mkozic yes please prove or disprove your hypothesis." }, { "id": 550959, "postDate": "2019-06-12T07:18:18.750Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 550853, "author_name": "DDgg", "author_url": "", "post_date": "2019-06-12T04:54:45.943000", "content": "Not sure why people downvote this (-4), its a personal hypothesis,\nmaybe true or false. It may not make sense to the reader or expert,\nit maybe crazy or the ideas or assumptions are completely wrong,\nhowever the only task is to prove or disprove it.
\n\nWe have a large blinded validation set, nobody knows \nthe answer (yet) so @mkozic just needs to prove it or disprove it.\nNo need to downvote unless some additional advice is given.
\n\nSOMs (Kohonen neural networks) were the holy grail of algorithms that\ncould solve everything in the 90s. Reading old papers and books\nthey literally could be used for everything (world peace, no hunger & healing cancer)
\n\nSee Google Scholar links for SOMs: \nhttps://scholar.google.be/scholar?hl=nl-be&q=som+Kohonen+neural+network+chemistry
\n\nSo @mkozic yes please prove or disprove your hypothesis.
", "votes": 0, "replies": [] }, { "id": 550959, "author_name": "", "author_url": "", "post_date": "2019-06-12T07:18:18.750000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "540941": "We could train a SOM on all the training data, including the data we have **only** for the train set. After that we could make a simplified SOM with the data that we have for both the training and test set, and train it only on train set. Then we could give that SOM test data and see where it places the train data, and with that info we could see where the test data would go in the SOM of all features. We could then train a regression network using the coordinates of train data in the SOM of all features (including other data), and use that info for predicting the constant of the test set. Would this be worth the effort?\n\nWhat I'm hoping to achieve with this is to use all the data we have for training set, and since SOMs would cluster training data with test data, that would give a prediction as to what other features for test set are, that we don't have.", "550853": "Not sure why people downvote this (-4), its a personal hypothesis,\nmaybe true or false. It may not make sense to the reader or expert,\nit maybe crazy or the ideas or assumptions are completely wrong,\nhowever the only task is to prove or disprove it.\n\nWe have a large blinded validation set, nobody knows \nthe answer (yet) so @mkozic just needs to prove it or disprove it.\nNo need to downvote unless some additional advice is given.\n\nSOMs (Kohonen neural networks) were the holy grail of algorithms that\ncould solve everything in the 90s. Reading old papers and books\nthey literally could be used for everything (world peace, no hunger & healing cancer)\n\nSee Google Scholar links for SOMs: \nhttps://scholar.google.be/scholar?hl=nl-be&q=som+Kohonen+neural+network+chemistry\n\nSo @mkozic yes please prove or disprove your hypothesis.", "550959": "" } }