Skip to main content

Infer.NET user guide : Tutorials and examples : Latent Dirichlet Allocation

LDA implementation

Page 1 | Page 2 | Page 3


This page describes a simple, reference implementation of LDA (with fixed Dirichlet priors) in Infer.NET. For background information, refer to the main page for the LDA example. The implementation described on this page assumes that the full graphical model along with all the inference messages can be stored in memory at once. The large number of these messages coupled with their large size means that only moderately sized problems can be solved with this implementation. This is partially alleviated by an inference option that supports sparse representations of these messages, and this allows much larger size problems to be handled in cases where topic distributions over words are truly expected to be sparse. For an implementation which also scales to support large data sets, refer to the shared variable version of LDA.

Both LDAModel and LDAShared are classes in the LDA project in src/Examples/LDA. The next section highlights aspects of the model definition in the LDAModel class. Page 2 shows inference and prediction capabilities. Page 3 describes the LDAShared class.

LDA Model Constructor

The model is defined in the constructor method for the LDAModel class which has the following arguments:

Argument Type Description Examples
sizeVocab int The size of the vocabulary 10000
numTopics int The number of latent topics 20

LDA Definition

First, a Range is defined for each plate in the graphical model. For words in documents, note that different documents will have different numbers of observed words, so in that case we need a jagged Range WInD. The NumWordsInDoc array is specified at inference time.

Range D = new Range(NumDocuments).Named("D");  
Range W = new Range(SizeVocab).Named("W");  
Range T = new Range(NumTopics).Named("T");  
NumWordsInDoc = Variable.Array<int>(D).Named("NumWordsInDoc");  
Range WInD = new Range(NumWordsInDoc[D]).Named("WInD");

Once we have defined the ranges, we can define the variables in the model. The main variables are the Theta (the mixture parameters for the per-document distributions over topics), and the Phi (the mixture parameters for the per-topic distributions over words); these are drawn respectively from Dirichlet priors ThetaPrior and PhiPrior which are set at run time (when training, these are set by specifying an alpha and beta which are the parameters of symmetric Dirichlet priors). The word observations are in the form of arrays of arrays of integer variables (indices into the vocabulary for each document). An additional jagged array variable gives the corresponding word counts. Indices and counts are observed at inference time.

Theta = Variable.Array<Vector>(D);  
ThetaPrior = Variable.Array<Dirichlet>(D).Named("ThetaPrior");  
Theta[D] = Variable<Vector>.Random(ThetaPrior[D]);  
Phi = Variable.Array<Vector>(T);  
PhiPrior = Variable.Array<Dirichlet>(T).Named("PhiPrior");  
Phi[T] = Variable<Vector>.Random(PhiPrior[T]);  
Words = Variable.Array(Variable.Array<int>(WInD), D).Named("Words");  
WordCounts = Variable.Array(Variable.Array<double>(WInD), D).Named("WordCounts");

The remainder of the model, defines the generative process whereby each observed word in a document is considered as a sample from a topic’s word distribution, where the topic is a sample from the document’s topic distribution, as described on the main LDA page.

using (Variable.ForEach(D))  
    using (Variable.ForEach(WInD))  
        using (Variable.Repeat(WordCounts[D][WInD]))  
            var topic = Variable.Discrete(Theta[D]);  
            using (Variable.Switch(topic))  
                Words[D][WInD] = Variable.Discrete(Phi[topic]);  

This completes the main definition of the model, but there are a couple of other things we need to set up before doing inference. Firstly note that we need to break symmetry when we’re doing inference because the model is symmetric with respect to the latent topics. We will do this by initialising the messages out of Theta. We will do this in a general way by defining a variable ThetaInit which we will observe at inference time to be a random message in the form of a Dirichlet distribution. The advantage of making ThetaInit a variable is that we can modify it without causing a model recompilation.

ThetaInit = Variable.Array<Dirichlet>(D).Named("ThetaInit");

Lastly, we would like to retrieve the evidence from the model so that we can compare models - for example we might want to optimise the alpha and beta hyper-parameters. We can compute model evidence by creating a mixture of the model with an empty model as follows:

var Evidence = Variable.Bernoulli(0.5).Named("Evidence");  
VariableIfBlock evidenceBlock = Variable.If(Evidence);  
    main model code  

Go to Page 2 for inference and prediction usage. Go to Page 3 for a scalable shared variable version of this model.

Page 1 | Page 2 | Page 3