MFF: Meta Fuzzy Functions for Aggregating Prediction Pools

MFF aggregates aligned numeric candidate predictions with weights obtained by clustering candidate behavior on validation data. Candidates may be trained inside R, varied across hyperparameters, or imported from other software.

library(MFF)
pool <- mff_example()
fit <- tune_mff(pool, clusters = 2:3, method = "fcm", m = c(1.5, 2),
                nstart = 3, reltol = 1e-6, control = mff_control(seed = 123))
summary(fit)
weights(fit)
pred <- predict(fit, pool)
evaluate(pred, test_observed(pool))
plot(fit, type = "weight_heatmap")

The example pool holds validation and test predictions of diamond prices from seven regression learners on a random split of ggplot2’s diamonds (?mff_diamonds).

Use mff_pool() for external validation/test matrices, model_train() for selected regression learners, and boot_train() for bootstrap candidates of one learner. mff_learner() wraps any regression method for both helpers. Without a separate validation set, model_train(…, resampling = mff_cv()) uses out-of-fold predictions of the training rows as validation predictions. as_mff_pool() converts caret model lists, tidymodels tuning results, stacks data stacks and fable forecasts into pools. rolling_train() builds forecasting pools by rolling-origin evaluation of user forecasting functions. Forecasting pools require ordered indexes and correctly generated external point forecasts. MFF validates alignment; users remain responsible for external training without information leakage.

Classes: mff_prediction_pool, mff_fit and mff_tuned (which extends mff_fit). Public accessors make examples independent of internal list layout.

See the prediction-pool, forecasting and migration vignettes. Version 1.0.0 replaces the 0.2.x interface; the migration vignette maps old calls to new ones.

Use nstart to control independent clustering starts. reltol is a relative objective tolerance for FCM and an absolute convergence threshold for PFCM and GK; see ?fit_mff for method-specific semantics. NULL keeps the defaults. PFCM follows Pal, Pal, Keller and Bezdek (2005) and GK the improved covariance estimation of Babuska, van der Veen and Kaymak (2002); both are implemented in the package. K-means does not support a non-NULL reltol.

mirror server hosted at Truenetwork, Russian Federation.