深度学习>> mixdir>> 返回
项目作者： const-ae

项目描述 ：
Cluster high dimensional categorical datasets
高级语言： R
项目地址: git://github.com/const-ae/mixdir.git
创建时间： 2018-01-20T19:17:24Z
项目社区：https://github.com/const-ae/mixdir
开源协议：
下载


mixdir

The goal of mixdir is to cluster high dimensional categorical datasets.

It can

  • handle missing data
  • infer a reasonable number of latent class (try mixdir(select_latent=TRUE))
  • cluster datasets with more than 70,000 observations and 60 features
  • propagate uncertainty and produce a soft clustering

A detailed description of the algorithm and the features of the package can be found in the the accompanying paper. If you find the package useful please cite

C. Ahlmann-Eltze and C. Yau, “MixDir: Scalable Bayesian Clustering for High-Dimensional Categorical Data”, 2018 IEEE 5th International Conference on Data Science and Advanced Analytics (DSAA), Turin, Italy, 2018, pp. 526-539.

Installation

  1. install.packages("mixdir")
  3. # Or to get the latest version from github
  4. devtools::install_github("const-ae/mixdir")

Example

Clustering the mushroom data set.

  1. # Loading the library and the data
  2. library(mixdir)
  3. set.seed(1)
  5. data("mushroom")
  6. # High dimensional dataset: 8124 mushroom and 23 different features
  7. mushroom[1:10, 1:5]
  8. #>    bruises cap-color cap-shape cap-surface    edible
  9. #> 1  bruises     brown    convex      smooth poisonous
  10. #> 2  bruises    yellow    convex      smooth    edible
  11. #> 3  bruises     white      bell      smooth    edible
  12. #> 4  bruises     white    convex       scaly poisonous
  13. #> 5       no      gray    convex      smooth    edible
  14. #> 6  bruises    yellow    convex       scaly    edible
  15. #> 7  bruises     white      bell      smooth    edible
  16. #> 8  bruises     white      bell       scaly    edible
  17. #> 9  bruises     white    convex       scaly poisonous
  18. #> 10 bruises    yellow      bell      smooth    edible

Calling the clustering function mixdir on a subset of the data:

  1. # Clustering into 3 latent classes
  2. result <- mixdir(mushroom[1:1000,  1:5], n_latent=3)

Analyzing the result

  1. # Latent class of of first 10 mushrooms
  2. head(result$pred_class, n=10)
  3. #>  [1] 3 1 1 3 2 1 1 1 3 1
  5. # Soft Clustering for first 10 mushrooms
  6. head(result$class_prob, n=10)
  7. #>               [,1]         [,2]         [,3]
  8. #>  [1,] 3.103495e-07 1.055098e-05 9.999891e-01
  9. #>  [2,] 9.998594e-01 4.683764e-06 1.359291e-04
  10. #>  [3,] 9.998944e-01 3.111462e-06 1.025194e-04
  11. #>  [4,] 5.778033e-04 7.114603e-08 9.994221e-01
  12. #>  [5,] 3.662625e-07 9.999992e-01 4.183025e-07
  13. #>  [6,] 9.996461e-01 8.764031e-08 3.537838e-04
  14. #>  [7,] 9.998944e-01 3.111462e-06 1.025194e-04
  15. #>  [8,] 9.997331e-01 5.822320e-08 2.668420e-04
  16. #>  [9,] 5.778033e-04 7.114603e-08 9.994221e-01
  17. #> [10,] 9.999999e-01 5.850067e-09 9.845112e-08
  18. pheatmap::pheatmap(result$class_prob, cluster_cols=FALSE,
  19.                   labels_col = paste("Class", 1:3))

  2. # Structure of latent class 1
  3. # (bruises, cap color either yellow or white, edible etc.)
  4. purrr::map(result$category_prob, 1)
  5. #> $bruises
  6. #>      bruises           no 
  7. #> 0.9998223256 0.0001776744 
  8. #> 
  9. #> $`cap-color`
  10. #>        brown         gray          red        white       yellow 
  11. #> 0.0001775934 0.0001819672 0.0001776373 0.4079822666 0.5914805356 
  12. #> 
  13. #> $`cap-shape`
  14. #>      bell    convex      flat    sunken 
  15. #> 0.3926736 0.4767291 0.1304197 0.0001776 
  16. #> 
  17. #> $`cap-surface`
  18. #>   fibrous     scaly    smooth 
  19. #> 0.0568571 0.4871396 0.4560033 
  20. #> 
  21. #> $edible
  22. #>       edible    poisonous 
  23. #> 0.9998223174 0.0001776826
  25. # The most predicitive features for each class
  26. find_predictive_features(result, top_n=3)
  27. #>       column    answer class probability
  28. #> 19 cap-color    yellow     1   0.9993990
  29. #> 22 cap-shape      bell     1   0.9990947
  30. #> 1    bruises   bruises     1   0.7089533
  31. #> 48    edible poisonous     3   0.9980468
  32. #> 15 cap-color       red     3   0.8462032
  33. #> 9  cap-color     brown     3   0.6473043
  34. #> 5    bruises        no     2   0.9990364
  35. #> 11 cap-color      gray     2   0.9978218
  36. #> 32 cap-shape    sunken     2   0.9936162
  37. # For example: if all I know about a mushroom is that it has a
  38. # yellow cap, then I am 99% certain that it will be in class 1
  39. predict(result, c(`cap-color`="yellow"))
  40. #>          [,1]         [,2]         [,3]
  41. #> [1,] 0.999399 0.0003004692 0.0003004907
  43. # Note the most predictive features are different from the most typical ones
  44. find_typical_features(result, top_n=3)
  45. #>         column  answer class probability
  46. #> 1      bruises bruises     1   0.9998223
  47. #> 43      edible  edible     1   0.9998223
  48. #> 19   cap-color  yellow     1   0.5914805
  49. #> 3      bruises bruises     3   0.9995546
  50. #> 27   cap-shape  convex     3   0.7460615
  51. #> 9    cap-color   brown     3   0.6746224
  52. #> 44      edible  edible     2   0.9995310
  53. #> 5      bruises      no     2   0.9713177
  54. #> 35 cap-surface fibrous     2   0.7355413

Dimensionality Reduction

  1. # Defining Features
  2. def_feat <- find_defining_features(result, mushroom[1:1000,  1:5], n_features = 3)
  3. print(def_feat)
  4. #> $features
  5. #> [1] "cap-color" "bruises"   "edible"   
  6. #> 
  7. #> $quality
  8. #> [1] 74.35146
  10. # Plotting the most important features gives an immediate impression
  11. # how the cluster differ
  12. plot_features(def_feat$features, result$category_prob)
  13. #> Loading required namespace: ggplot2
  14. #> Loading required namespace: tidyr

Underlying Model

The package implements a variational inference algorithm to solve a Bayesian latent class model (LCM).