machine learning - Training on imbalanced data using TensorFlow -


the situation:

i wondering how use tensorflow optimally when training data imbalanced in label distribution between 2 labels. instance, suppose mnist tutorial simplified distinguish between 1's , 0's, images available either 1's or 0's. straightforward train using provided tensorflow tutorials when have 50% of each type of image train , test on. case 90% of images available in our data 0's , 10% 1's? observe in case, tensorflow routinely predicts entire test set 0's, achieving accuracy of meaningless 90%.

one strategy have used success pick random batches training have distribution of 0's , 1's. approach ensures can still use of training data , produced decent results, less 90% accuracy, more useful classifier. since accuracy useless me in case, metric of choice typically area under roc curve (auroc), , produces result respectably higher .50.

questions:

(1) strategy have described accepted or optimal way of training on imbalanced data, or there 1 might work better?

(2) since accuracy metric not useful in case of imbalanced data, there metric can maximized altering cost function? can calculate auroc post-training, can train in such way maximize auroc?

(3) there other alteration can make cost function improve results imbalanced data? currently, using default suggestion given in tensorflow tutorials:

cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(pred, y)) optimizer = tf.train.adamoptimizer(learning_rate=learning_rate).minimize(cost)

i have heard may possible up-weighting cost of miscategorizing smaller label class, unsure of how this.

i'm 1 struggling imbalanced data. strategy counter imbalanced data below.

1) use cost function calculating 0 , 1 labels @ same time below.

cost = tf.reduce_mean(-tf.reduce_sum(y*tf.log(_pred) + (1-y)*tf.log(1-_pred), reduction_indices=1))

2) use smote, oversampling method making number of 0 , 1 labels similar. refer here, http://comments.gmane.org/gmane.comp.python.scikit-learn/5278

both strategy worked when tried make credit rating model.

logistic regression typical method handle imbalanced data , binary classification such predicting default rate. auroc 1 of best metric counter imbalanced data.


Comments

Post a Comment

Popular posts from this blog

authentication - Mongodb revoke acccess to connect test database -

i have ec2 instance mongodb 3.2.0 installed. have user admin access. can connect user , perform operation remotely. 1 can connect database test remote if knows ip mongo 11.11.11.11. although not able make operations, want restrict access user credentials can connect server. first need create users , grant roles users, example: use admin db.createuser( { user: "myuseradmin", pwd: "abc123", roles: [ { role: "useradminanydatabase", db: "admin" } ] } ) then need start server --auth parameter mongod --auth more information here
Read more

r - Update two sets of radiobuttons reactively - shiny -

i have read ( how make choices in radiobuttons reactive in shiny? ) shows me how update radiobutton s in reactive way. however, when try , update 2 sets of buttons same data, 1 set renders. example: server: # create example data wafer <- rep(c(1:3), each=3) length <- c(1,1,2,1,1,1,3,5,1) width <- c(3,1,6,1,1,1,1,1,6) dd <- data.frame(wafer, length, width) shinyserver(function(input, output, session){ # create reactive dataframe store data values <- reactivevalues() values$df <- data.frame() # lengths , widths of wafer user input <- eventreactive(input$do, { subset(dd, wafer %in% input$wafer, select = length:width) }) # update reactive data frame lengths , widths have been selected user input observe({ if(!is.null(a())) { values$df <- rbind(isolate(values$df), a()) } }) output$wl <- renderdatatable({ a() }) # update radio buttons unique length , widths stored in values$df # ever "observe" put first in co...
Read more

ios - Realm over CoreData should I use NSFetchedResultController or a Dictionary? -

i'm working on app using core data , nsfetchedresultcontroller many reasons switch use realm. saw there "realm version" of nsfetchedresultcontroller on github wouldn't compatible current code using core data. the view controller displaying list of people same school, address book. list sublist of people studied in same city. so thinking make 1 unique request database retrieve list of people , filter locally list within 1 dictionary per school [string: anyobject] string section name within tableview , anyobject array of people. first solution, make nsfetchedresultcontroller each school predicate filtering location. delete actions etc.. handled delegates -> not compatible realm create dictionaries , update them each actions... -> works realm it's annoying code. any better solution? edit: i need clarify request: i'd write class inherit uitableviewcontroller. this class has list of people sorted in alphabetical order the tableview has ...
Read more