Neural network model variables

[WonderHog]

Since I only have a high-level understanding of neural networks this may be a dumb question, but is there a way to get at the output variables and coefficients from the neural network tool?  I would like to have something more than just a black box to report to my down-stream users.

Thanks.

[AndrewL]

Hi @WonderHog,

There is a Model Coefficients macro in the Predictive District on the Analytics Gallery (gallery.alteryx.com). Not sure if it works for the Neural Network or not. Other than that I'm not aware of any way to pull coefficients from the predictive tools aside from possibly some custom R code.

[SydneyF · Accepted solution]

Hi @WonderHog

Sorry for the delay in this post - I was drafting a response and then it unintentionally fell off of my radar.

Neural Networks don't have variable coefficients like traditional regression models. The closest analog to coefficients in a Neural Network are the weights that connect variables in a neural network. The weights can be used to describe relationships between variables and estimate variable importance. Like @AndrewL mentioned, there is a way to extract some model details with custom R code. If you have an R coding background, one option would be to check out the R package NeuralNetTools, which is compatible with nnet (The R package the Alteryx Neural Network Tool uses) and provides visualization and analysis tools to aid in interpreting Neural Networks. You could bring these functions into Alteryx with the R tool.

For ways to assess variables with the Alteryx Neural Network Tool's native abilities, the (optional) Effects Plots in the Report (R) output of the Neural Network Tool can be used to interpret the effect of the variables on different classifications. These plots depict how the probability of a record being categorized as a given classification changes as the predictor variable changes, so you can visualize the relationship between the predictor variable and the target, averaging over the effect of other predictor fields.

Hopefully this helps!

SydneyF - Customer Support Engineer

[TimothyL]

Hi @WonderHog ,

We recently published a new model explanation macros: LIME here which supports Neural Network. 

More than that, it supports other predictive tools including SVM, Decision Tree, Random Forest and more. Give a try and hope it could give you a hand