errors "Error in Ops.Surv(observed, predicted) : Invalid operation on a survival time"

[MosquitoFan]

I used the mlr3 and mlr3proba with surv.xgboost as follows, since xgboost cannot output survival matrix so I used a pipeline, `xgb_task = TaskSurv$new("train", backend = data.frame(train), time = "time", event = "event")

xgb_lrn = as_learner(ppl(c("distrcompositor"), lrn("surv.xgboost", objective = "survival:cox", nrounds=300L, eta=0.1), form = "ph")) `

then I try to explain the model, I set the predict_survival_function and predict_cumulative_hazard_function, which could be successfully conducted when I run them, explainer <- explain(model = xgb_lrn, data = train, y = Surv(time[!fold == i1],event[!fold == i1]), predict_function = function(model,newdata){ predict(model, newdata, predict_type = "<Prediction>")$crank }, predict_survival_function = function(model,newdata,times){ t(predict(model, newdata, predict_type = "<Prediction>")$distr$survival(times)) }, predict_cumulative_hazard_function = function(model,newdata,times){ t(predict(model, newdata, predict_type = "<Prediction>")$distr$cumHazard(times)) } )

Preparation of a new explainer is initiated -> model label : R6 ( default ) -> data : 1498 rows 62 cols -> target variable : 1498 values -> predict function : function(model, newdata) { predict(model, newdata, predict_type = "")$crank } -> predicted values : No value for predict function target column. ( default ) -> model_info : package mlr3 , ver. 0.14.0 , task regression ( default ) -> predicted values : numerical, min = -7.026592 , mean = -0.8300239 , max = 11.02893
-> residual function : difference between y and yhat ( default ) -> residuals : the residual_function returns an error when executed ( WARNING )

• I could successfully run predict(explainer,train) to get the risk score

However, I meet error when adding output_type

Error in predict_function(model, newdata, ...) : unused argument (output_type = "chf")

• I also meet error when try to explain model_parts(explainer)

Error in Ops.Surv(observed, predicted) : Invalid operation on a survival time

model_performance(explainer)

Error in Ops.Surv(y, predict_function(model, data)) : Invalid operation on a survival time

I don't know what's wrong with my code.

[mikolajsp]

Hello @MosquitoFan,

I think the issue is that the wrong kind of explainer is created (from the DALEX package) as making the pipeline doesn't give the correct class to the pipeline object. (We would expect it to be LearnerSurv).

However, the possible fix is simple:

• replace explain() function with explain_survival() when creating your explainer.
• So the code should look like this:

  explainer <- explain_survival(model = xgb_lrn,  data = train, y = Surv(time[!fold == i1],event[!fold == i1]), predict_function = function(model,newdata){ predict(model, newdata, predict_type = "<Prediction>")$crank }, predict_survival_function = function(model,newdata,times){ t(predict(model, newdata, predict_type = "<Prediction>")$distr$survival(times)) }, predict_cumulative_hazard_function = function(model,newdata,times){ t(predict(model, newdata, predict_type = "<Prediction>")$distr$cumHazard(times)) } )

I've tested it on some of my data and it works. Let me know if it solved your issue :)

P.S. If you see residuals ... in the last line of output when creating an explainer you're creating a DALEX not a survex explainer.

[MosquitoFan]

Hi mikolaj,

Thank you so much, it works and the plot looks beautiful!