From: A model-agnostic approach for understanding heart failure risk factors
Step number | Steps |
---|---|
1 | fit a machine learning model (for a neural network model, the activation function of the output layer needs to be a sigmoid function) |
2 | determine a reference value \({x}_{i}^{r}\) |
3 | \(\left.{es}_{i,j}=logit\left(f\left({x}_{1}^{j},\dots ,{x}_{i}^{j},\dots ,{x}_{n}^{j}\right)\right)-logit(f({x}_{1}^{j},\dots ,{x}_{i}^{r},\dots ,{x}_{n}^{j})\right)\) - where \(f(.)\) is the prediction of the probability of the positive class by the model - if \({x}_{i,j}={x}_{i,k}\), then consider the average of them |
4 | \({ES}_{i}= \sum_{j=1}^{n}|{es}_{i,j}|\) |
5 | For a continuous feature, plot \({es}_{i,j}\) against the value of \({x}_{i, j \epsilon n \left\{obs\right\}}\) to depict the effect of \({x}_{i}\) at different values on the output with respect to the reference |
6 | Rank \({ES}_{i}\) of categorial and continuous features to compare strength of different features |