deepforest: Deep Forest
In michaelzxu/deepforest:

Description Usage Arguments Value

View source: R/deepforest.R

A combination of stacking (per Kaggle) and the deep forest idea proposed in https://arxiv.org/pdf/1702.08835.pdf with a few additional knobs and levers of my own..

deepforest(x, y, x_val = NULL, y_val = NULL, nfold = 5, index = NULL,
  objective = NULL, eval_func = NULL, nlayer = 5, nmeta = 4,
  nmetarep = 1, nclass = NULL, metatype = rep(c("bag", "boost", "dart",
  "lin"), length = nmeta * nmetarep), metaparam = NULL, metarandom = FALSE,
  colsample_bylayer = 1, colsample_add = FALSE, accumulate = FALSE,
  nthread = 2, missing = NA, printby = "fold", ...)

`x`	Object coercible to a matrix.
`y`	A vector of labels for `x`.
`x_val`	Object coercible as matrix used for validation. Defaults to `x` if not specified.
`y_val`	A vector of labels for `x_val`.
`nfold`	The number of folds to stack on. If `index` is specified, this argument is ignored.
`index`	A list of integers indicating observations belonging to each fold. Random sampling is used if `NULL` to create `nfold` folds.
`objective`	The objective passed to `xgb.train`. It is inferred by nclass if not specified.
`eval_func`	A function used to evaluate each meta model.
`nlayer`	The number of layers to stack.
`nmeta`	The number of meta models to build on each layer.
`nmetarep`	The number of repeated meta models to construct.
`nclass`	An integer indicating the number of classes. `1` for regression, `2` for binary classification and `>2` for multiclass. If `NULL`, then it is set to `length(unique(y))` if the value is less than 10 and 1 otherwise.
`metatype`	A list of character variables indicating type of meta models to construct. Currently supports "bag", "boost", "dart" and "lin" for a random forest, boosted trees, boosted trees with dropout and boosted linear models respectively, all based on xgboost. If length of `metatype` is not equal to `nmeta * nmetarep` then it is recycled to that length.
`metaparam`	A list indicating additional parameters for each meta model. The mechanism used to generate default values and create randomized values are in `metaparam`. If list is shorter than `nmeta * nmetarep` then all meta models without a clear `metaparam` will take on default values, so you should indicate manually tuned models first in `metatype`.
`metarandom`	A logical indicating whether to create randomized miscellaneous parameters.
`colsample_bylayer`	A numeric vector between [0,1] indicating the fraction of features to sample at each layer. If a single value, then you may use `colsample_add` to further modify how the value changes, otherwise a constant value will be used for each layer.
`colsample_add`	Either a logical or a numeric. If `TRUE` then values will be added at each layer to ensure that the final `nlayer` has a sampling rate of 1 (i.e. all features used). If `FALSE` then no modifier is applied to `colsample_bylayer`. If a numeric value, that constant value is added to `colsample_bylayer` after each layer.
`accumulate`	A logical indicating if the predictions by meta models from each layer is accumulated. If `FALSE` then each layer will not have the direct predictions from two layers before although it will be implicitly approximated by the predictions from the layer immediately before.
`nthread`	The nthread argument passed to `xgb.train`.
`missing`	The `missing` argument passed to `xgb.train`.
`...`	Additional arguments passed to `xgb.train`.