All new and revised torch functions no longer require the “cito” package.
Added new argument algo = c(“nodewise”,“layerwise”,“structured”,“neuralgraph”) to the SEMdnn() function. Four algorithms are now implemented using R MLPs (number of nodes with non-zero incoming connectivity) for “nodewise”, L<R MLPs (number of layers in the input graph) for “layerwise”, and 1 MLP for “structured” or “neuralgraph”.
Delete algo=“nn” and algo=“dnn” in the SEMml() function. These NNs for a small neural network model (1 hidden layer and 10 nodes) and large neural network model (1 hidden layers and 1000 nodes) run with algo=“nodewise” of the SEMdnn() function.
Added new function getLOCO(). For SEMml() algo=c(“sem”,“tree”,“rf”,“xgb”), computes the contributions of each variable to individual predictions using LOCO (Leave Out COvariates) values based on ghost variables. A CPU-efficient procedure.
Various fixed bugs discovered after the release 1.0.0.
Version 1.0.0 is a major release with several new features, including:
Added new argument outcome = NULL (defult). This parameter is used in SEMdnn() and SEMml() functions to process a sink categorical node (as a factor) for classification purposes using all graph nodes as covariates.
Added new argument newoutcome = NULL (defult). This parameter is used in predict (.SEM, .DNN, .ML) functions to predict a sink categorical node (as a factor) for classification purposes using all graph nodes as covariates.
classificationReport() function. A report showing the main classification metrics, like precision, recall, F1-score, accuracy, Matthew’s correlation coefficient (mcc) for all classes of the node = as.factor(outcome).
crossValidation() function. A R-repeated K-fold cross-validation with a list of M models from SEMrun(), SEMml() and SEMdnn(). The winning model is selected by reporting the mean predicted performances across all RxKxM runs.
getVariableImportance() function. Extraction of common Machine Learning (ML) variable (predictor) importance measures after fitting SEMrun(), SEMml() or SEMdnn() models.
Added new argument nboot = 0 (default). This parameter implements cheap bootstrapping in SEMdnn() and SEMml() functions to generate uncertainties, i.e. CIs, for DNN/ML parameters. Bootstrapping can be enabled by setting a small number (from 1 to 10) of bootstrap samples.
Change argument thr = 0.5 * max(abs(parameters)) (default). Now the DAG can be colored using a numeric [0-1] threshold. For example, 1/0.5 = 2, can be interpreted as the number of times a node/edge parameter is less than the maximum parameter value.
Various fixed bugs discovered after the release 0.1.0.