Predict documents with a fitted HyperGAT classifier — predict.hg

Reuses the trained network, vocabulary, preprocessing and topic keywords. Words absent from the fitted vocabulary are ignored. A document with no remaining words receives missing predictions and a warning, preserving its row and id. The winning softmax probability is a model score, not a calibrated probability of correctness.

Usage

# S3 method for class 'hg_hypergat'
predict(
  object,
  newdata = NULL,
  column = NULL,
  id = NULL,
  labels = NULL,
  type = c("predictions", "probabilities"),
  ...
)

Arguments

object: A fitted hg_hypergat.
newdata: A character vector or data frame of new documents. NULL returns the original predictions.
column, id: Text and id columns of a data frame. Defaults to the columns used when fitting. An unnamed character vector receives sequential document ids.
labels: Optional observed labels for newdata, in the same formats as hg_hypergat(). Supplying them evaluates the predictions through hg_get.hg_classification() and plot(). These labels never update the model. Unscorable labelled documents count as errors.
type: "predictions" (default), a document table, or "probabilities", a matrix with one column per fitted class.
...: Unused.

Value

A plain data frame (node, label, predicted, score, margin) or a probability matrix, in input order. No fitting occurs.

References

Ding, K., Wang, J., Li, J., Li, D., & Liu, H. (2020). Be more with less: Hypergraph attention networks for inductive text classification. EMNLP 2020, 4927-4936.