hawkrobe / group-inference Goto Github PK

make the topic model

• 1. keep working on posterior predictive check

  • * some information criterion, likelihood measure
  • * graphing data over predicted distribution

• 2. interpretability

  • * transitions/combination that most distinguish the clusters
  • * some divergence mesuares, KS test maybe

• try to get smoother K-metric curve by ↑ nsteps in sparse-learning and subsetting input/prediction set
• inference performance as a function of N observed transitions, make computationally feasible by subsetting

• debug multi_obs model (expand out assignment)
• search for indices (i.e. transitions) that are maximally informative about group

• explore the relation between number of groups/topics and modeling results (f(K))
• compare the posterior predictions between group and topic models
• interactions between the two items above: do the groups and the topics correspond between the two models? do they diverge when K gets larger? rationale being the two models can be seen as soft/hard versions of the same thing, so maybe they will just pick up similar structure anways.
  • robert's hypothesis is that they will diverge
• add stickbreaking to topic model

• is there more variability among people's target-specific predictions than their generic predictions
• hierarchical models that allow for specific partner level variations? (socio-linguisticky)
• scoping - specific to the concept of emotion predicition, or something about generc-specific dynamics in social cognition?

• try on other data
  • e.g. ratings for close others vs. self-ratings
  • normalizing vs. not normalizing the rows of the transitions
  • diagonal vs. no diagonal
• double-check model output using MCMC against the variational inference (and running with multiple random seeds)
• Look at predictions of mixture model to see how well it's fitting, and examine residuals
• Making more clustering (UMAP/tsne) visualizations

• refactoring the code, separate out the models into a script, separate out data preprocessing, and maybe save the preprocessed data as objects
• finish up topic model for normalized data
• the two should give the same thing when K = 1
• when K = 2 how much do the groups and topics correspond