1. Mixed membership model

This model wants to discover a set of memberships

In contrast, cluster models aim at discovering a single membership

In clustering:

• one topic indicator z_i per document i
• all words come from(get scored under) same topic z_i
• distribution on prevaluence of topics in corpus,  πi=[π_i1 ... π_ik]

In LDA:

• one topic indicator z_iw per word in doc i
• each word gets socred under its topic  z_iw 
• distribution on prevaluence of topics in document, πi=[π_i1 ... π_ik]

LDA inputs: set of words per doc for each doc in corpus

LDA outputs: corpus-wide topic vocab distributions, topic assignments per word, topic proportions per doc

Typically LDA is specified as a bayesian model

• account for unvertainty in parameters when making predictions
• naturally regularizes parameter estimates in contrast to MLE.

 

2. Gibbs sampling

Iterative random hard assignments

predictions:

• make prediction for each snapshot of randomly assigned variables/parameters
• average predictions for final result
• look at snapshot of randomly assigned variables/parameters that maximize joint model probability

benefits:

• intuitive updates
• very straightforward to implement

Procedure:

• randomly reassign all z_iw based on doc topic proportions and topic vocab distributions
• randomly reassign doc topic proportions based on assignments z_iw in current doc
• repeat for all docs
• randomly ressign topic vocab distributions based on assignments z_iw in entire corpus
• repeat steps 1-4 until max iter reached

3. Collapsed gibbs sampling

Based no special structure of LDA model, can sample just indicator variables z_iw.

no need to sample other parameters

• corpus-wide topic vocab distributions
• per-doc topic proportions

Procedure:

randomly reassign z_iw based on current assignment z_jv of all other words in document and corpus.

How much doc likes each topic based on other assignments in doc

 

 

 

n_ik is the current assignment to topic k in doc i

N_i is the words in doc i

α is the smoothing param from bayes prior

 

How much each topic likes the word dynamic based on assignments in other docs in corpus

m_dynamic,k is the assignments corpus-wide of word dynamic to topic k

γ is the smoothing param

V is the size of vocab

 

probabilities = how much doc likes topic * how much topic likes word(normalize this product of terms over k possible topics)

Based on the probabilities increment count based on new assignmentof z_iw

what to do with the collapsed samples?

From best sample of z_{iw, can infer}

• Topics from conditional distribution
• document embedding