Clustering of mixed data

This page provides access to material for benchmarking « ready-to-use » clustering methods for mixed data. It refers to a paper recently published in Nature Scientific Reports and entitled

« Head-to-head comparison of clustering methods for heterogeneous data. A simulation-driven benchmark »

by

Gregoire Preud’homme ^{1, 2}, Kevin Duarte ¹, Kevin Dalleau ³, Claire Lacomblez ¹, Emmanuel Bresso ³, Malika Smaïl-Tabbone ^2,3,  Miguel Couceiro ³, Marie-Dominique Devignes ^2,3, Masatake Kobayashi ^{1, 2}, Olivier Huttin ^{1, 2}, João Pedro Ferreira ^{1, 2}, Faiez Zannad ^{1, 2}, Patrick Rossignol ^{1, 2}, Nicolas Girerd ^{1, 2}

1. Université de Lorraine, Centre d’Investigations Cliniques Plurithématique 1433, INSERM 1116, CHRU de Nancy, France.
2. F-CRIN INI-CRCT Cardiovascular and Renal Clinical Trialists Network, France.
3. Université de Lorraine, CNRS, Inria Nancy Grand-Est, LORIA, UMR 7503, Vandoeuvre-lès-Nancy, France.

---

Simulated datasets

For each tested design 1000 simulated datasets were generated (see paper for the method used). The sets of 1000 datasets can be downloaded as a single rds file (formatted for usage with R packages) named according to the design parameters.
Default parameters are as follows

• Population size : 300 ;
• Number of clusters : 6 ;
• Number of continuous variables : 4 ;
• Degree of relevance of continuous variables : mild ;
• Proportion of relevant continuous variables : 100% (for a total of 4 variables);
• Total number of categorical variables : 4 ;
• Degree of relevance of categorical variables : mild.
• Proportion of relevant categorical variables : 100% (for a total of 4 variables);

In the table, for each design, the values of the varying parameters are indicated with a link for downloading the corresponding dataset (average size 10Mo).

---

Tested algorithms

Distance-based

• • Gower dissimilarity : cluster package (daisy function).
  • Unsupervised Extra Tree dissimilarity : UET package, yet unpublished, available at gitlab address, use build_randomized_tree_and_get_sim function.
  • Partitioning Around Medoids (PAM) : clustMixType package (PAM function).
  • Hierarchical Ascendant Clustering : stats (R-base package, hclust function).
  • K-Prototypes : clustMixType package (kproto function).

Model-based

• • Kamila : kamila package (kamila function).
  • Latent Class Analysis (LCA) : poLCA package (poLCA function).
  • Latent Class Model (LCM) : VarSelLCM package (VarSelCluster function).
  • Clustering by Mixture Modeling (MixMod) : Rmixmod package (mixmodCluster function).

---