---
title: "Speeding up computations"
author: "Marjolein Fokkema"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Faster computation}
  %\VignetteEngine{knitr::rmarkdown}
  %\usepackage[utf8]{inputenc}
bibliography: bib.bib
csl: bib_style.csl
---

The computational load of function `pre` can be heavy. This vignette shows some ways to reduce it. 

There are two main steps in fitting a rule ensemble: 1) Rule generation and 2) Estimation of the final ensemble. Both can be adjusted to reduce computational load. 


## Faster rule generation

### Tree fitting algorithm

By default, `pre` uses the conditional inference tree algorithm [@HothyHorn06] as implemented in function `ctree` of `R` package `partykit` [@HothyZeil15] for rule induction. The main reason is that it does not present with a selection bias towards variables with a greater number of possible cut points. Yet, use of `ctree` brings a relatively heavy computational load:

```{r}
airq <- airquality[complete.cases(airquality), ]
airq$Month <- factor(airq$Month)
library("pre")
set.seed(42)
system.time(airq.ens <- pre(Ozone ~ ., data = airq))
summary(airq.ens)
```

Computational load can be substantially reduced by employing the CART algorithm of [@BreiyFrie84] as implemented in function `rpart` from the package of the same name by @TheryAtki22. This can be specified through the `tree.unbiased` argument:

```{r}
set.seed(42)
system.time(airq.ens.cart <- pre(Ozone ~ ., data = airq, tree.unbiased = FALSE))
summary(airq.ens.cart)
```

Alternatively, rules can be generated using the random-forest approach originally proposed by [@brei01] as implemented in function `randomForest` from the package of the same name by @LiawyWien02: 

```{r}
set.seed(42)
system.time(airq.ens.rf <- pre(Ozone ~ ., data = airq, randomForest = TRUE))
summary(airq.ens.rf)
```

Note, however, that the resulting ensembles will likely be more complex and also present with a selection bias towards variables with a greater number of possible cutpoints. The higher complexity is also observed above, where CART and random forest resulted in a substantially larger number of terms. This is due to the default stopping criteria for `rpart` and `randomForest` being considerably less conservative than that of `ctree`. This will result in the generation of more and longer rules, which will also tend to increase complexity of the final ensemble. Furthermore, those algorithms prefer to split using variables with a larger number of cutpoints, and this bias may propagate to the final rule ensemble.

```{r,echo=FALSE, eval=FALSE}
sort(sapply(airq, \(x) length(unique(x))), decr = TRUE)
par(mfrow = c(1, 3))
importance(airq.ens, cex.lab = .7, cex.axis = .7, cex.main = .7, 
           main = "Variable importances (ctree)")
importance(airq.ens.cart, cex.lab = .7, cex.axis = .7, cex.main = .7, 
           main = "Variable importances (CART)")
imps.rf <- importance(airq.ens.rf, cex.lab = .7, cex.axis = .7, cex.main = .7, 
                      main = "Variable importances (randomForest)")
```

### Maximum rule depth

Reducing tree (and thereby rule) depth will reduce computational load of both rule fitting and estimation of the final ensemble: 

```{r}
set.seed(42)
system.time(airq.ens.md <- pre(Ozone ~ ., data = airq, maxdepth = 1L))
summary(airq.ens.md)
```

Likely, reducing the maximum depth will improve interpretability, but may decrease predictive accuracy for the final ensemble.



### Number of trees

By default, 500 trees are generated. Computation time can be reduced substantially by reducing the number of trees. This may of course negatively impact predictive accuracy. When using a smaller number of trees, it is likely beneficial to increase the learning rate (`learnrate = .01`, by default) accordingly: 

```{r}
set.seed(42)
system.time(airq.ens.nt <- pre(Ozone ~ ., data = airq, ntrees = 100L, learnrate = .05))
summary(airq.ens.nt)
```

## Faster estimation of the final ensemble

Function `cv.glmnet` from package `glmnet` is used for fitting the final model. The relevant arguments of function `cv.glmnet` can be passed directly to function `pre`.


### Parallel computation

For example, parallel computation can be employed by specifying `par.final = TRUE` in the call to `pre` (and registering parallel beforehand, e.g., using `doMC`). Parallel computation will not affect performance of the final ensemble. Note that parallel computation will only reduce computation time for datasets with a (very) large number of observations. For smaller datasets, the use of parallel computation may even increase computation time. 



### Number of cross-validation folds

The number of cross-validation repetitions can be also be reduced, but this will probably not reduce computation time much and may negatively affect predictive performance of the final ensemble.

```{r}
set.seed(42)
system.time(airq.ens.nf <- pre(Ozone ~ ., data = airq, nfolds = 5L))
summary(airq.ens.nf)
```

## References