I wrote a fast version of Naive Bayes that requires pypy for the base function then I wrapped it for R use (with data.table). I'd like to post it here to see if anyone is using NB and a code check. The probabilities aren't normalized, but I think the code is correct (it may not be). It's fairly fast and should work on the full train data and test data in minutes.
\n\nR does not have a good Naive Bayes. They all seem to memory leak or freeze on me. My function requires the input train data.frame to have the first column as target (0...numclasses-1) and the other columns as features (use correct header names), and the test data.frame to have the features with correct header names.
", "messages": [ { "id": "119299", "postDate": "05/08/2016 23:58:05", "content": "I wrote a fast version of Naive Bayes that requires pypy for the base function then I wrapped it for R use (with data.table). I'd like to post it here to see if anyone is using NB and a code check. The probabilities aren't normalized, but I think the code is correct (it may not be). It's fairly fast and should work on the full train data and test data in minutes.
\n\nR does not have a good Naive Bayes. They all seem to memory leak or freeze on me. My function requires the input train data.frame to have the first column as target (0...numclasses-1) and the other columns as features (use correct header names), and the test data.frame to have the features with correct header names.
", "rawMarkdown": "I wrote a fast version of Naive Bayes that requires pypy for the base function then I wrapped it for R use (with data.table). I'd like to post it here to see if anyone is using NB and a code check. The probabilities aren't normalized, but I think the code is correct (it may not be). It's fairly fast and should work on the full train data and test data in minutes.\r\n\r\nR does not have a good Naive Bayes. They all seem to memory leak or freeze on me. My function requires the input train data.frame to have the first column as target (0...numclasses-1) and the other columns as features (use correct header names), and the test data.frame to have the features with correct header names.", "votes": null }, { "id": "119309", "postDate": "05/09/2016 04:36:49", "content": "My first scoring model is one I deemed hierarchical, though in fact it looks a lot like some of the Python scripts that have been posted. However, my count-and-sort had some particular nuances. That got me to 0.487 or so. In the process, it occurred to me just how much what I had come up with was like NB.
\n\nI am using the e1071 package in R, the naiveBayes function (different from NaiveBayes) and, frankly, I think it flies. However, even varying the feature selection and using Laplace smoothing at l=1, 2, and 3, I haven't yet submitted any worthwhile scores. I tinkered with eps and threshold in naiveBayes, but they didn't result in any train/validation improvements either. I'm due for another round of trying NB, and on my 32GB laptop, I'm not hesitant to work in R Studio with the e1071 implementation. In fact, I did build models on the entire train data set!
", "rawMarkdown": "My first scoring model is one I deemed hierarchical, though in fact it looks a lot like some of the Python scripts that have been posted. However, my count-and-sort had some particular nuances. That got me to 0.487 or so. In the process, it occurred to me just how much what I had come up with was like NB.\r\n\r\nI am using the e1071 package in R, the naiveBayes function (different from NaiveBayes) and, frankly, I think it flies. However, even varying the feature selection and using Laplace smoothing at l=1, 2, and 3, I haven't yet submitted any worthwhile scores. I tinkered with eps and threshold in naiveBayes, but they didn't result in any train/validation improvements either. I'm due for another round of trying NB, and on my 32GB laptop, I'm not hesitant to work in R Studio with the e1071 implementation. In fact, I did build models on the entire train data set!", "votes": null }, { "id": "119353", "postDate": "05/09/2016 12:41:26", "content": "Thats really cool Mike, thanks for sharing ! I will try it out. (p.s. good to learn how R and python can interact)
", "rawMarkdown": "Thats really cool Mike, thanks for sharing ! I will try it out. (p.s. good to learn how R and python can interact)", "votes": null }, { "id": "121036", "postDate": "05/23/2016 05:45:43", "content": "I have only gotten ~42.5 LB with partial data, but 30.8 MAP local validation.
\n\n", "rawMarkdown": "I have only gotten ~42.5 LB with partial data, but 30.8 MAP local validation.\r\n\r\nhttps://www.kaggle.com/yukatherin/expedia-hotel-recommendations/naive-bayes-with-countvectorizer/run/244962", "votes": null } ], "comments": [ { "id": 119309, "author_name": "siliconvalley", "author_url": "", "post_date": "05/09/2016 04:36:49", "content": "My first scoring model is one I deemed hierarchical, though in fact it looks a lot like some of the Python scripts that have been posted. However, my count-and-sort had some particular nuances. That got me to 0.487 or so. In the process, it occurred to me just how much what I had come up with was like NB.
\n\nI am using the e1071 package in R, the naiveBayes function (different from NaiveBayes) and, frankly, I think it flies. However, even varying the feature selection and using Laplace smoothing at l=1, 2, and 3, I haven't yet submitted any worthwhile scores. I tinkered with eps and threshold in naiveBayes, but they didn't result in any train/validation improvements either. I'm due for another round of trying NB, and on my 32GB laptop, I'm not hesitant to work in R Studio with the e1071 implementation. In fact, I did build models on the entire train data set!
", "votes": null, "replies": [] }, { "id": 119353, "author_name": "darraghdog", "author_url": "", "post_date": "05/09/2016 12:41:26", "content": "Thats really cool Mike, thanks for sharing ! I will try it out. (p.s. good to learn how R and python can interact)
", "votes": null, "replies": [] }, { "id": 121036, "author_name": "yukatherin", "author_url": "", "post_date": "05/23/2016 05:45:43", "content": "I have only gotten ~42.5 LB with partial data, but 30.8 MAP local validation.
\n\n", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "119299": "I wrote a fast version of Naive Bayes that requires pypy for the base function then I wrapped it for R use (with data.table). I'd like to post it here to see if anyone is using NB and a code check. The probabilities aren't normalized, but I think the code is correct (it may not be). It's fairly fast and should work on the full train data and test data in minutes.\r\n\r\nR does not have a good Naive Bayes. They all seem to memory leak or freeze on me. My function requires the input train data.frame to have the first column as target (0...numclasses-1) and the other columns as features (use correct header names), and the test data.frame to have the features with correct header names.", "119309": "My first scoring model is one I deemed hierarchical, though in fact it looks a lot like some of the Python scripts that have been posted. However, my count-and-sort had some particular nuances. That got me to 0.487 or so. In the process, it occurred to me just how much what I had come up with was like NB.\r\n\r\nI am using the e1071 package in R, the naiveBayes function (different from NaiveBayes) and, frankly, I think it flies. However, even varying the feature selection and using Laplace smoothing at l=1, 2, and 3, I haven't yet submitted any worthwhile scores. I tinkered with eps and threshold in naiveBayes, but they didn't result in any train/validation improvements either. I'm due for another round of trying NB, and on my 32GB laptop, I'm not hesitant to work in R Studio with the e1071 implementation. In fact, I did build models on the entire train data set!", "119353": "Thats really cool Mike, thanks for sharing ! I will try it out. (p.s. good to learn how R and python can interact)", "121036": "I have only gotten ~42.5 LB with partial data, but 30.8 MAP local validation.\r\n\r\nhttps://www.kaggle.com/yukatherin/expedia-hotel-recommendations/naive-bayes-with-countvectorizer/run/244962" }, "source": "meta" }