tag:blogger.com,1999:blog-9145120678290195131.post567697637168805981..comments2024-03-21T16:22:06.514+02:00Comments on Probably Approximately a Scientific Blog: Supervised LearningVered Shwartzhttp://www.blogger.com/profile/17531957962535846245noreply@blogger.comBlogger6125tag:blogger.com,1999:blog-9145120678290195131.post-49389859520617805612019-01-24T21:49:26.061+02:002019-01-24T21:49:26.061+02:00This comment has been removed by a blog administrator.Dr Purva Piushttps://www.blogger.com/profile/05883980841903455890noreply@blogger.comtag:blogger.com,1999:blog-9145120678290195131.post-52613732107376256552019-01-24T21:49:01.964+02:002019-01-24T21:49:01.964+02:00This comment has been removed by a blog administrator.Dr Purva Piushttps://www.blogger.com/profile/05883980841903455890noreply@blogger.comtag:blogger.com,1999:blog-9145120678290195131.post-91895246315034628332015-08-14T18:59:25.679+03:002015-08-14T18:59:25.679+03:00First of all, it's just as an aid to the docto...First of all, it's just as an aid to the doctor - no machine learning algorithm is accurate enough to diagnose a patient and prescribe a drug :)<br /><br />As for your question, some models can handle missing values (and for others, I assume there are heuristics about how to complete them, but I never experimented with that).Vered Shwartzhttps://www.blogger.com/profile/17531957962535846245noreply@blogger.comtag:blogger.com,1999:blog-9145120678290195131.post-74618229271381531672015-08-14T18:44:38.316+03:002015-08-14T18:44:38.316+03:00"Medical diagnosis - predict whether a patien..."Medical diagnosis - predict whether a patient suffers from a certain disease, based on his symptoms"<br /><br />What if they don't run all tests? Is there machine learning where some items don't have information regarding some features?shwartzhttps://www.blogger.com/profile/02744914397813357107noreply@blogger.comtag:blogger.com,1999:blog-9145120678290195131.post-57928868635759538372015-08-11T11:16:18.212+03:002015-08-11T11:16:18.212+03:00You hit it very hard :)
It depends on the specifi...You hit it very hard :)<br /><br />It depends on the specific algorithm (there are several such algorithm to compute a binary model). In general, these algorithms start with some random model and try to improve it with every training instance they see. A model is usually a weight given to every feature according to its indicativeness of the class (e.g. the word "cash" is highly indicative of spam, so its weight will be high).<br /><br />For every training instance, if the current model predicts its label incorrectly (different from the true label), the algorithm will change the model a bit in the "direction" of the current training instance (so that the model could predict it "more correctly"). Every change is applied to all the features (to strengthen the feature values of the current instance). <br /><br />Overfitting is caused when the model considers all the features of the training instances as indicative, even though some of them are non-indicative (specific to some training instances, and not generally to most negative / positive instances). Applying regularization means reducing the score of a certain model that has too many indicative features. By penalizing the model this way, it would make it only increase the weights of features that occurred in many training instances (of the same class) -- the indicative features.<br /><br />I hope it's not too complicated :)Vered Shwartzhttps://www.blogger.com/profile/17531957962535846245noreply@blogger.comtag:blogger.com,1999:blog-9145120678290195131.post-53596808998855764662015-08-11T10:45:23.471+03:002015-08-11T10:45:23.471+03:00So how do you punish the algorithm?So how do you punish the algorithm?Shirinoreply@blogger.com