utiml
Utilities for Multi-Label Learning
Multi-label learning strategies and others procedures to support multi-
label classification in R. The package provides a set of multi-label procedures such as
sampling methods, transformation strategies, threshold functions, pre-processing
techniques and evaluation metrics. A complete overview of the matter can be seen in
Zhang, M. and Zhou, Z. (2014)
The utiml package is a framework to support multi-label processing, like Mulan on Weka.
The main methods available on this package are organized in the groups:
- Classification methods
- Evaluation methods
- Pre-process utilities
- Sampling methods
- Threshold methods
Instalation
The installation process is similar to other packages available on CRAN:
install.packages("utiml")This will also install mldr. To run the examples in this document, you also need to install the packages:
# Base classifiers (SVM and Random Forest)install.packages(c("e1071", "randomForest"))Install via github (development version)
devtools::install_github("rivolli/utiml")Multi-label Classification
Running Binary Relevance Method
library(utiml) # Create two partitions (train and test) of toyml multi-label datasetds <- create_holdout_partition(toyml, c(train=0.65, test=0.35)) # Create a Binary Relevance Model using e1071::svm methodbrmodel <- br(ds$train, "SVM", seed=123) # Predictprediction <- predict(brmodel, ds$test) # Show the predictionshead(as.bipartition(prediction))head(as.ranking(prediction)) # Apply a thresholdnewpred <- rcut_threshold(prediction, 2) # Evaluate the modelsresult <- multilabel_evaluate(ds$tes, prediction, "bipartition")thresres <- multilabel_evaluate(ds$tes, newpred, "bipartition") # Print the resultprint(round(cbind(Default=result, RCUT=thresres), 3))Running Ensemble of Classifier Chains
library(utiml) # Create three partitions (train, val, test) of emotions datasetpartitions <- c(train = 0.6, val = 0.2, test = 0.2)ds <- create_holdout_partition(emotions, partitions, method="iterative") # Create an Ensemble of Classifier Chains using Random Forest (randomForest package)eccmodel <- ecc(ds$train, "RF", m=3, cores=parallel::detectCores(), seed=123) # Predictval <- predict(eccmodel, ds$val, cores=parallel::detectCores())test <- predict(eccmodel, ds$test, cores=parallel::detectCores()) # Apply a thresholdthresholds <- scut_threshold(val, ds$val, cores=parallel::detectCores())new.val <- fixed_threshold(val, thresholds)new.test <- fixed_threshold(test, thresholds) # Evaluate the modelsmeasures <- c("subset-accuracy", "F1", "hamming-loss", "macro-based") result <- cbind( Test = multilabel_evaluate(ds$tes, test, measures), TestWithThreshold = multilabel_evaluate(ds$tes, new.test, measures), Validation = multilabel_evaluate(ds$val, val, measures), ValidationWithThreshold = multilabel_evaluate(ds$val, new.val, measures)) print(round(result, 3))More examples and details are available on functions documentations and vignettes, please refer to the documentation.
News
Changelog
utiml 0.1.4
New Features
- MLKNN algorithm
- ranking-loss baseline
- label problem evaluation measures
- kfold bult-in method
- The foodtruck dataset
- ESL algorithm
Minor changes
- confusion matrix in matrix format
Bug fixes
- Stratification sampling to support instances without labels
- Fixed threshold with multiple values
- Update documentation
utiml 0.1.3
Major changes
- Change
multilabel_evaluationto also return the label measures
Bug fixes
- Bugfix in
brplusbecause the newfeatures were using different levels - Fix
baselineusing hamming-loss to prevent empty label prediction - Fix empty prediction when all labels have the same probability
Minor changes
- Fix type mistakes in documentation
utiml 0.1.2
Major changes
- change base.method parameter name for base.algorithm
Bug fixes
- Bugfix in
homerto deal with labels without intances and to predict instances based on the meta-label scores - Refactory of merge_mlconfmat
- Ensure reproducibility in all cases
utiml 0.1.1
New multi-label transformation methods including pairwise and multiclass approaches. Some fixes from previous version.
Major changes
- lcard threshold calibration
- Use categorical attributes in multilabel datasets and methods
- LIFT multi-label classification method
- RPC multi-label classification method
- CRL multi-label classification method
- LP multi-label classification method
- RAkEL multi-label classification method
- BASELINE multi-label classification method
- PPT multi-label classification method
- PS multi-label classification method
- EPS multi-label classification method
- HOMER multi-label classification method
Minor changes
- Add Empty Model as base method to fix training labels with few examples
multilabel_confusion_matrixaccepts a data.frame or matrix with the predicitons- Change EBR and ECC to use threshold calibration
- Include empty.prediction configuration to enable/disable empty predictions
Bug fixes
- Majority Ensemble Predictions Votes
- Majority Ensemble Predictions Probability
- Base method not found message error
- Base method support any attribute names
- Normalize data ignore attributes with a single value
- MBR support labels without positive examples
- Fix average precision and coverage measures to support instances without labels
utiml 0.1.0
First release of utiml:
- Classification methods:
Binary Relevance (BR);BR+;Classifier Chains;ConTRolled Label correlation exploitation (CTRL);Dependent Binary Relevance (DBR);Ensemble of Binary Relevance (EBR);Ensemble of Classifier Chains (ECC);Meta-Binary Relevance (MBR or 2BR);Nested Stacking (NS);Pruned and Confident Stacking Approach (Prudent); and,Recursive Dependent Binary Relevance (RDBR) - Evaluation methods: Create a multi-label confusion matrix and multi-label measures
- Pre-process utilities: fill sparse data; normalize data; remove attributes; remove labels; remove skewness labels; remove unique attributes; remove unlabeled instances; and, replace nominal attributes
- Sampling methods: Create subsets of multi-label dataset; create holdout and k-fold partitions; and, stratification methods
- Threshold methods: Fixed threshold; MCUT; PCUT; RCUT; SCUT; and, subset correction
- Synthetic dataset: toyml