= Topic Model =

A '''topic model''' is an application of machine learning to qualitative methods. Topics of documents are extracted and optionally assigned to a bin.

<<TableOfContents>>

----



== Data Structure ==

Topic models make use of a '''Document-Term matrix''' ('''DTM'''). Rows are ''documents'' and columns are ''tokens''.



----



== Implementation ==

Common topic model methods include:

 * '''Probabilistic Latent Semantic Analysis''' ('''PLSA''')
 * '''Latent Dirichlet Allocation''' ('''LDA''')
 * '''Pachinko Allocation'''

LDA is a form of unsupervised learning. It takes a few inputs as priors. Importantly, κ is the number of topics. With a higher α (between 0 and 1), more topics will be extracted from each document. Short documents likely call for a small α value. With a higher β (between 0 and 1), topics are composed of more words. Dense documents likely call for a high β value.

Topic models are evaluated by '''perplexixity''' of the model, '''coherence''' of topics, and '''exclusivity''' of the topics.

See `textmineR::FitLdaModel`.

LSA steps are:

 1. Run models with varying α and β values.
    * a.k.a. '''hyperparameter tuning'''
    * Using the median likely value of κ, run a model for each combination of discrete likely values of α and β
    * Minimize for perplexity and take the optimal pair.
    * e.g. `text2vec::perplexity`
 2. Run models with varying κ values.
    * Using the optimal (α, β) pair, run a model for each likely κ value.
 3. Select the κ value that gives the best trade-off of minimizing perplexity, maximizing coherence, and maximizing exclusivity.
    * e.g. `text2vec::perplexity`, `textmineR::CalcPropCoherence`, and `topicdoc::topic_exclusivity`
 4. Examine the topics.
    * e.g. `textmineR::SummarizeTopics`
    * γ terms are the most exclusive tokens for each topic.
    * φ terms are the most common tokens for each topic.
    * If these sets of tokens reveal that the modeled topics are insufficient, go back and select a different κ value.



----
CategoryRicottone