This is the web implementation for Durkon, a library for automatically building transparent predictive models.

You can use Prepare Model to create a starting model, use Train Model to fit it to a dataset, and use Predict With Model to apply it to new data (you will need to copypaste the model between subpages).

Check the hovertextLike this! if you need more information.

Processing is all done client-side, in-browser: information provided stays on your computer the entire time. This also means you can model while disconnected from the internet.


NOTE: The (far more powerful) Python library this is based off is here; a demo showing how to get started using it is here.

DataCopypaste your dataset here. Make sure it's in csv format!: Width,Height,Icing Thickness,Flavour,Wedding,Fancy,Price 13,7,6,vanilla,No,Somewhat,63.5 8,6,9,lemon,No,Somewhat,13.26 13,6,6,vanilla,No,Somewhat,88.26 6,6,4,marzipan,No,Somewhat,8.76 8,10,5,chocolate,Yes,Somewhat,24.95 7,8,10,lemon,No,Somewhat,22.06 9,6,7,chocolate,No,Very,15.52


Categorical featuresPredictive features which are expressed as categories: things like "Species", "Genre", or "Manufacturer_of_Car". List them with a comma between each feature name, and no spaces.:

Continuous featuresPredictive features which are expressed as quantities: things like "Height", "Page_Count", or "Age_of_Driver". List them with a comma between each feature name, and no spaces.:

Response variableThe quantity you're trying to predict. (In an Insurance context, this is usually something like "Price" or "Cost".):

Weight column The importance assigned to each row; leave this blank to treat all rows equally. (In an Insurance context, this is usually where you'd put Exposure when modelling Frequency or Burning Cost.):

Prep

ModelThe model you want to train. You can get an untrained model of the right format from Prepare Model.:

DataCopypaste your dataset here. Make sure it's in csv format!: Width,Height,Icing Thickness,Flavour,Wedding,Fancy,Price 13,7,6,vanilla,No,Somewhat,63.5 8,6,9,lemon,No,Somewhat,13.26 13,6,6,vanilla,No,Somewhat,88.26 6,6,4,marzipan,No,Somewhat,8.76 8,10,5,chocolate,Yes,Somewhat,24.95 7,8,10,lemon,No,Somewhat,22.06 9,6,7,chocolate,No,Very,15.52


Response variableThe quantity you're trying to predict. (In an Insurance context, this is usually something like "Price" or "Cost".):

Weight column The importance assigned to each row; leave this blank to treat all rows equally. (In an Insurance context, this is usually where you'd put Exposure when modelling Frequency or Burning Cost.):


Training roundsThe number of iterations you want the training algorithm to run for. More rounds are almost always better, but A) more rounds take longer to run, B) returns to adding more rounds diminish sharply, and C) overtraining makes it more likely that you'll overfit to the training data.:

Learning rateThe rate at which the model is trained. If training takes too long, try increasing this value; if you get crashes or other odd behaviour, try decreasing it. Note that the optimal value here will vary for different loss functions.:

Loss functionThe error distribution the model uses to decide which mis-predictions are most important to correct. Convention is to use Gamma when predicting the price of something, Poisson when predicting how many times something will happen in a given timespan, and Tweedie for situations which are a fusion of Gamma and Poisson (such as predicting Burning Cost in an Insurance context).: Gamma Poisson Tweedie

Tweedie ExponentThe 'p' parameter for the Tweedie loss function; only used when Tweedie is selected. Takes a value between 1 and 2: 1 is Poisson, 2 is Gamma, and somewhere inbetween is somewhere inbetween.:

Train

ModelThe model you want to predict with. You can get an untrained model of the right format from Prepare Model, and train it using Train Model.:

Predict For Single Row Predict For Single Row (With Graphs) Predict For Dataset