Skip to contents

Choose the number of topics by coherence and exclusivity

Source: R/text_topic_model.R
hg_topic_search.Rd

Fits hg_topics() for every number of topics in k and scores each model by the mean semantic coherence (UMass; Mimno et al. 2011) and the mean exclusivity (FREX; Bischof & Airoldi 2012) of its topics, computed by hg_topic_quality() on the n most probable words. This is the diagnostic of stm::searchK() (Roberts, Stewart & Tingley 2019). Adding topics usually makes them more exclusive and less coherent, so no single number is best; Roberts et al. (2014) choose among the models on the frontier of the two, those that no other model beats on both. A model is on the frontier when no other number of topics has both a higher mean coherence and a higher mean exclusivity.

Usage

hg_topic_search(
  hg,
  k,
  nstart = 1L,
  n = 10L,
  frexw = 0.7,
  max_iter = 1000L,
  tol = 1e-05,
  seed = 1L,
  parallel = FALSE,
  n_cores = 2L
)

# S3 method for class 'hypernets_topic_search'
plot(x, ...)

Arguments

hg

A net_hg or text_hypergraph.

k

Numbers of topics to compare, at least two whole numbers.

nstart

Number of random starts (default 10).

n

Top words per topic scored (default 10).

frexw

Weight on exclusivity in FREX, in (0, 1) (default 0.7).

max_iter

Maximum multiplicative updates per start (default 1000).

tol

Relative change of the divergence between two checks, ten updates apart, below which a start has converged (default 1e-5); 0 runs every start for max_iter updates.

seed

Seed of the first start; start r uses seed + r - 1. The caller's random-number stream is restored on exit. Default 1.

parallel

Fit the numbers of topics with parallel::mclapply() (not on Windows); every fit seeds itself, so the result equals the serial one. Default FALSE.

n_cores

Cores when parallel = TRUE (default 2).

x

A table returned by the verb.

...

Unused.

Value

A base data.frame of class hypernets_topic_search, one row per number of topics: k, coherence and exclusivity (means over the topics), divergence (of the kept start), agreement (mean agreement of the topics across the starts; NA with nstart = 1), converged and frontier. plot() draws exclusivity against coherence, one labelled point per number of topics, the frontier joined by a line and drawn as filled points. Conditions are those of hg_topics(); a start that does not converge raises hypernets_no_converge.

References

Roberts, M. E., Stewart, B. M., & Tingley, D. (2019). stm: An R package for structural topic models. Journal of Statistical Software, 91(2), 1-40. doi:10.18637/jss.v091.i02

Roberts, M. E., Stewart, B. M., Tingley, D., Lucas, C., Leder-Luis, J., Gadarian, S. K., Albertson, B., & Rand, D. G. (2014). Structural topic models for open-ended survey responses. American Journal of Political Science, 58(4), 1064-1082. doi:10.1111/ajps.12103

Mimno, D., Wallach, H. M., Talley, E., Leenders, M., & McCallum, A. (2011). Optimizing semantic coherence in topic models. Proceedings of EMNLP 2011, 262-272.

Bischof, J. M., & Airoldi, E. M. (2012). Summarizing topical content with word frequency and exclusivity. Proceedings of ICML 2012, 201-208.

See also

hg_topics(), hg_topic_quality().

Examples

corpus <- c(
  a = "soup salt onion soup broth", b = "salt soup broth onion",
  c = "stars sky moon night", d = "sky stars night moon moon",
  e = "soup stars salt sky night broth", f = "onion salt stars broth")
search <- hg_topic_search(text_hypergraph(corpus), k = 2:3, nstart = 2,
                          n = 4)
search
#>   k  coherence exclusivity divergence agreement converged frontier
#> 1 2 -0.4887538    3.240670   4.241503 0.5833333      TRUE     TRUE
#> 2 3 -1.3415043    3.242266   1.927450 0.6888889      TRUE     TRUE