R/predict-methods.r

In tsensembler: Dynamic Ensembles for Time Series Forecasting

#' @title Predicting new observations using an ensemble
#'
#' @name predict
#' @aliases predict.ade predict.dets predict.base
#' @rdname predict-methods
#'
#' @description Initially, the predictions of the base models are collected.
#' Then, the predictions of the loss to be incurred by the base models \strong{E_hat}
#' (estimated by their associate meta models) are computed. The weights of
#' the base models are then estimated according to \strong{E_hat} and the committee of
#' top models. The committee is built according to the \emph{lambda} and
#' \emph{omega} parameters. Finally, the predictions are combined
#' according to the weights and the committee setup.
#'
#' @param object an object of class \code{\link{ADE-class}};
#' @param newdata new data to predict
#'
#'
#' @export
NULL

#' @rdname predict-methods
#'
#' @examples
#'
#' ###### Predicting with an ADE ensemble
#'
#' specs <- model_specs(
#'  learner = c("bm_glm", "bm_mars"),
#'  learner_pars = NULL
#' )
#'
#' data("water_consumption")
#' dataset <- embed_timeseries(water_consumption, 5)
#' train <- dataset[1:1000, ]
#' test <- dataset[1001:1500, ]
#'
#' model <- ADE(target ~., train, specs)
#'
#' preds <- predict(model, test)
#'
#'
#' @export
setMethod("predict",
          signature("ADE"),
          function(object, newdata) {
            data_size <- nrow(newdata)
            seq. <- seq_len(data_size)
            N <- object@base_ensemble@N

            Y_hat <- predict(object@base_ensemble, newdata)
            Y_hat_recent <- predict(object@base_ensemble, object@recent_series)
            Y_hat_recent <- utils::tail(Y_hat_recent, object@lambda)

            Y_hat_extended <- rbind.data.frame(Y_hat_recent, Y_hat)

            Y <- get_y(newdata, object@form)

            E_hat <-
              lapply(object@meta_model,
                     function(o) {
                       meta_predict(o, newdata, object@meta_model_type)
                     })
            names(E_hat) <- colnames(Y_hat)

            E_hat <- abs(as.data.frame(E_hat))

            Y_rs <- get_y(object@recent_series, object@form)
            Y_rs <- utils::tail(Y_rs, object@lambda)

            if (!object@all_models) {
              Y_rs <- get_y(object@recent_series, object@form)
              Y_rs <- utils::tail(Y_rs, object@lambda)

              C <-
                build_committee(
                  rbind.data.frame(Y_hat_recent, Y_hat),
                  c(Y_rs, Y),
                  lambda = object@lambda,
                  omega = object@omega)

              C <- C[-seq_len(object@lambda)]
              C <- lapply(C, unname)
            } else
              C <- NULL

            if (!object@all_models) {
              W <- matrix(0., ncol = N, nrow = data_size)
              for (j in seq.) {
                W_j <- E_hat[j, C[[j]]]
                W_j <- model_weighting(W_j, object@aggregation)

                W[j,  C[[j]]] <- W_j
                W[j, -C[[j]]] <- 0.
              }
              colnames(W) <- colnames(Y_hat)
            } else {
              W <- apply(E_hat, 1, model_weighting, trans = object@aggregation)
              W <- t(W)
            }

            if (object@select_best) {
              W <- select_best(W)
              C <- NULL
            }

            if (object@sequential_reweight) {
              ssimilarity <- sliding_similarity(Y_hat_extended, object@lambda)
              W <- sequential_reweighting(ssimilarity, W)
            }

            y_hat <-
              combine_predictions(
                Y_hat = Y_hat,
                W = W,
                committee = C
              )

            ade_hat(y_hat, Y_hat, C, E_hat)
          })


#' @rdname predict-methods
#'
#' @examples
#' \dontrun{
#'
#' ###### Predicting with a DETS ensemble
#'
#' specs <- model_specs(
#'  learner = c("bm_svr", "bm_glm", "bm_mars"),
#'  learner_pars = NULL
#' )
#'
#' data("water_consumption")
#' dataset <- embed_timeseries(water_consumption, 5)
#' train <- dataset[1:700, ]
#' test <- dataset[701:1000, ]
#'
#' model <- DETS(target ~., train, specs, lambda = 50, omega = .2)
#'
#' preds <- predict(model, test)
#' }
#'
#' @export
setMethod("predict",
          signature("DETS"),
          function(object, newdata) {
            seq. <- seq_len(nrow(newdata))

            Y_hat <- predict(object@base_ensemble, newdata)

            Y <- get_y(newdata, object@form)

            Y_hat_recent <-
              predict(object@base_ensemble, object@recent_series)
            Y_recent <- get_y(object@recent_series, object@form)

            scores <-
              model_recent_performance(rbind.data.frame(Y_hat_recent, Y_hat),
                                       c(Y_recent, Y),
                                       object@lambda,
                                       object@omega,
                                       object@base_ensemble@pre_weights)

            scores$top_models <- scores$top_models[-seq_len(object@lambda)]
            scores$model_scores <- scores$model_scores[-seq_len(object@lambda), ]

            if (object@select_best) {
              scores$model_scores <- select_best(scores$model_scores)
              C <- NULL
            }

            y_hat <-
              combine_predictions(
                Y_hat = Y_hat,
                W = scores$model_scores,
                committee = scores$top_models
              )

            dets_hat(
              y_hat = y_hat,
              Y_hat = Y_hat,
              Y_committee = scores$top_models,
              W = scores$model_scores
            )
          })


#' @rdname predict-methods
#'
#' @examples
#'
#' \dontrun{
#' ###### Predicting with a base ensemble
#'
#' model <- ADE(target ~., train, specs)
#'
#' basepreds <- predict(model@base_ensemble, test)
#' }
#'
#'
#' @export
setMethod("predict",
          signature("base_ensemble"),
          function(object, newdata) {
            seq. <- seq_len(object@N)

            Y_hat <-
              lapply(seq.,
                     function(o) {
                       compute_predictions(object@base_models[o],
                                           object@form,
                                           newdata)
                     })

            Times <- sapply(Y_hat, attr, "Times")

            Y_hat <- unlist(Y_hat, recursive = FALSE)
            nmodels <- names(Y_hat)

            Y_hat <- as.data.frame(Y_hat)
            colnames(Y_hat) <- nmodels

            attr(Y_hat, "Times") <- Times

            Y_hat
          })