v1.x compatible.
A Julia package for variational Bayesian topic modeling.
Topic models are Bayesian hierarchical models designed to discover the latent low-dimensional thematic structure within corpora. Topic models are fit using either Markov chain Monte Carlo (MCMC), or variational inference (VI).
Markov chain Monte Carlo methods are slow but consistent, given enough time MCMC will fit the exact model asymptotically. Contrarily, variational inference is fast but inconsistent, as one must approximate distributions in order to ensure tractability.
This package takes the latter approach to topic modeling.
(@v1.8) pkg> add https://github.com/ericproffitt/TopicModelsVB.jlDelimitedFiles
SpecialFunctions
LinearAlgebra
Random
Distributions
OpenCL
CrayonsIncluded in TopicModelsVB.jl are two datasets:
- National Science Foundation Abstracts 1989 - 2003:
- 128804 documents
- 25319 vocabulary
- CiteULike Science Article Database:
- 16980 documents
- 8000 vocabulary
- 5551 users
Let's begin with the Corpus data structure. The Corpus data structure has been designed for maximum ease-of-use. Datasets must still be cleaned and put into the appropriate format, but once a dataset is in the proper format and read into a corpus, it can easily be modified to meet the user's needs.
There are four plaintext files that make up a corpus:
- docfile
- vocabfile
- userfile
- titlefile
None of these files are mandatory to read a corpus, and in fact reading no files will result in an empty corpus. However in order to train a model a docfile will be necessary, since it contains all quantitative data known about the documents. On the other hand, the vocab, user and title files are used solely for interpreting output.
The docfile should be a plaintext file containing lines of delimited numerical values. Each document is a block of lines, the number of which depends on what information is known about the documents. Since a document is at its essence a list of terms, each document must contain at least one line containing a nonempty list of delimited positive integer values corresponding to the terms of which it is composed. Any further lines in a document block are optional, however if they are present they must be present for all documents and must come in the following order:
terms - A line of delimited positive integers corresponding to the terms which make up the document (this line is mandatory).
counts - A line of delimited positive integers, equal in length to terms, corresponding to the number of times a term appears in a document.
readers - A line of delimited positive integers corresponding to those users which have read the document.
ratings - A line of delimited positive integers, equal in length to readers, corresponding to the rating each reader gave the document.
An example of a single doc block from a docfile with all possible lines included,
...
4,10,3,100,57
1,1,2,1,3
1,9,10
1,1,5
...
The vocab and user files are tab delimited dictionaries mapping positive integers to terms and usernames (resp.). For example,
1 this
2 is
3 a
4 vocab
5 file
A userfile is identitcal to a vocabfile, except usernames will appear in place of vocabulary terms.
Finally, a titlefile is simply a list of titles, not a dictionary, and is of the form,
title1
title2
title3
title4
title5
The order of these titles correspond to the order of document blocks in the associated docfile.
To read a corpus into Julia, use the following function,
readcorp(;docfile="", vocabfile="", userfile="", titlefile="", delim=',', counts=false, readers=false, ratings=false)The file keyword arguments indicate the path where the respective file is located.
It is often the case that even once files are correctly formatted and read, the corpus will still contain formatting defects which prevent it from being loaded into a model. Therefore, before loading a corpus into a model, it is important that one of the following is run,
fixcorp!(corp)or
fixcorp!(corp, pad=true)Padding a corpus will ensure that any documents which contain vocab or user keys not in the vocab or user dictionaries are not removed. Instead, generic vocab and user keys will be added as necessary to the vocab and user dictionaries (resp.).
The fixcorp! function allows for significant customization of the corpus object.
For example, let's begin by loading the CiteULike corpus,
corp = readcorp(:citeu)A standard preprocessing step might involve removing stop words, removing terms which appear less than 200 times, and alphabetizing our corpus.
fixcorp!(corp, stop=true, abridge=200, alphabetize=true, trim=true)
## Generally you will also want to trim your corpus.
## Setting trim=true will remove leftover terms from the corpus vocabulary.After removing stop words and abridging our corpus, the vocabulary size has gone from 8000 to 1692.
A consequence of removing so many terms from our corpus is that some documents may now by empty. We can remove these documents from our corpus with the following command,
fixcorp!(corp, remove_empty_docs=true)In addition, if you would like to preserve term order in your documents, then you should refrain from condesing your corpus.
For example,
corp = Corpus(Document(1:9), vocab=split("the quick brown fox jumped over the lazy dog"))
showdocs(corp) ●●● Document 1
the quick brown fox jumped over the lazy dog
fixcorp!(corp, condense=true)
showdocs(corp) ●●● Document 1
jumped fox over the quick dog lazy brown the
Important. A corpus is only a container for documents.
Whenever you load a corpus into a model, a copy of that corpus is made, such that if you modify the original corpus at corpus-level (remove documents, re-order vocab keys, etc.), this will not affect any corpus attached to a model. However! Since corpora are containers for their documents, modifying an individual document will affect it in all corpora which contain it. Therefore,
-
Using
fixcorp!to modify the documents of a corpus will not result in corpus defects, but will cause them also to be changed in all other corpora which contain them. -
If you would like to make a copy of a corpus with independent documents, use
deepcopy(corp). -
Manually modifying documents is dangerous, and can result in corpus defects which cannot be fixed by
fixcorp!. It is advised that you don't do this without good reason.
The available models are as follows:
LDA(corp, K)
Latent Dirichlet allocation model with K topics.
fLDA(corp, K)
Filtered latent Dirichlet allocation model with K topics.
CTM(corp, K)
Correlated topic model with K topics.
fCTM(corp, K)
Filtered correlated topic model with K topics.
CTPF(corp, K)
Collaborative topic Poisson factorization model with K topics.gpuLDA(corp, K)
GPU accelerated latent Dirichlet allocation model with K topics.
gpuCTM(corp, K)
GPU accelerated correlated topic model with K topics.
gpuCTPF(corp, K)
GPU accelerated collaborative topic Poisson factorization model with K topics.Let's begin our tutorial with a simple latent Dirichlet allocation (LDA) model with 9 topics, trained on the first 5000 documents from the NSF corpus.
using TopicModelsVB
using Random
using Distributions
Random.seed!(7);
corp = readcorp(:nsf)
corp.docs = corp[1:5000];
fixcorp!(corp, trim=true)
## It's strongly recommended that you trim your corpus when reducing its size in order to remove excess vocabulary.
## Notice that the post-fix vocabulary is smaller after removing all but the first 5000 docs.
model = LDA(corp, 9)
train!(model, iter=150, tol=0)
## Setting tol=0 will ensure that all 150 iterations are completed.
## If you don't want to compute the ∆elbo, set checkelbo=Inf.
## training...
showtopics(model, cols=9, 20)topic 1 topic 2 topic 3 topic 4 topic 5 topic 6 topic 7 topic 8 topic 9
research system data theory research research research research plant
problems research earthquake study university data project study cell
design data project problems support project study chemistry species
systems systems research research students study data high protein
algorithms control study equations program ocean social studies cells
parallel time soil work science water understanding properties plants
data design damage investigator award studies economic chemical studies
project project seismic principal scientists processes important materials research
based analysis response project dr provide information structure genetic
models processing structures geometry sciences field policy program gene
model solar sites mathematical projects time development surface study
system computer ground systems conference important work reactions molecular
analysis information analysis differential scientific climate theory electron proteins
techniques high information algebraic national marine provide metal dna
methods techniques materials groups engineering models political experimental dr
problem development provide space provide measurements science molecular genes
performance models buildings analysis project sea models systems important
computer developed results methods year species change energy understanding
work based important solutions researchers understanding scientific project specific
developed image program finite mathematical global studies phase determine
If you are interested in the raw topic distributions. For LDA and CTM models, you may access them via the matrix,
model.beta
## K x V matrix
## K = number of topics.
## V = number of vocabulary terms, ordered identically to the keys in model.corp.vocab.Now that we've trained our LDA model we can, if we want, take a look at the topic proportions for individual documents.
For instance, document 1 has topic breakdown,
println(round.(topicdist(model, 1), digits=3))
## = [0.0, 0.0, 0.0, 0.0, 0.0, 0.435, 0.082, 0.0, 0.482]This vector of topic weights suggests that document 1 is mostly about biology, and in fact looking at the document text confirms this observation,
showdocs(model, 1)
## Could also have done showdocs(corp, 1). ●●● Document 1
●●● CRB: Genetic Diversity of Endangered Populations of Mysticete Whales: Mitochondrial DNA and Historical Demography
commercial exploitation past hundred years great extinction variation sizes
populations prior minimal population size current permit analyses effects
differing levels species distributions life history...
Just for fun, let's consider one more document (document 25),
println(round.(topicdist(model, 25), digits=3))
## = [0.0, 0.0, 0.0, 0.849, 0.0, 0.149, 0.0, 0.0, 0.0]
showdocs(model, 25) ●●● Document 25
●●● Mathematical Sciences: Nonlinear Partial Differential Equations from Hydrodynamics
work project continues mathematical research nonlinear elliptic problems arising perfect
fluid hydrodynamics emphasis analytical study propagation waves stratified media techniques
analysis partial differential equations form basis studies primary goals understand nature
internal presence vortex rings arise density stratification due salinity temperature...
We see that in this case document 25 appears to be about environmental computational fluid dynamics, which corresponds precisely to topics 4 and 6.
Furthermore, if we want to, we can also generate artificial corpora by using the gencorp function.
Generating artificial corpora will in turn run the underlying probabilistic graphical model as a generative process in order to produce entirely new collections of documents, let's try it out,
Random.seed!(7);
artificial_corp = gencorp(model, 5000, laplace_smooth=1e-5)
## The laplace_smooth argument governs the amount of Laplace smoothing (defaults to 0).
artificial_model = LDA(artificial_corp, 9)
train!(artificial_model, iter=150, tol=0, checkelbo=10)
## training...
showtopics(artificial_model, cols=9)topic 1 topic 2 topic 3 topic 4 topic 5 topic 6 topic 7 topic 8 topic 9
system plant research research research research project theory data
research species project design study university data study research
data cell study problems chemistry support earthquake problems project
systems studies data algorithms high students research research study
control protein social systems properties program structures equations water
project cells important parallel studies science study work ocean
models genetic economic project chemical award response geometry field
processing plants understanding data materials scientists soil investigator provide
high research policy models reactions sciences program principal important
analysis molecular information based program dr materials mathematical earthquake
solar dna development system phase scientific information project analysis
design gene work model structure projects structural differential effects
time proteins political analysis experimental engineering seismic algebraic studies
computer study provide methods surface national sites groups time
performance genes models techniques electron conference provide systems marine
For our next model, let's upgrade to a (filtered) correlated topic model (fCTM).
Filtering the correlated topic model will dynamically identify and suppress stop words which would otherwise clutter up the topic distribution output.
Random.seed!(7);
model = fCTM(corp, 9)
train!(model, tol=0, checkelbo=Inf)
## training...
showtopics(model, 20, cols=9)topic 1 topic 2 topic 3 topic 4 topic 5 topic 6 topic 7 topic 8 topic 9
algorithms earthquake theory students ocean economic chemistry physics protein
design data problems science water social chemical optical cell
parallel soil equations support sea theory metal solar cells
system damage geometry university climate policy reactions high plant
systems species investigator research marine political molecular laser species
performance seismic mathematical program measurements market surface particle gene
problems ground principal sciences data labor materials quantum genetic
network sites algebraic conference pacific decision organic devices proteins
networks response differential scientific global women molecules electron dna
control buildings space scientists atmospheric factors compounds materials plants
based forest groups national species human reaction radiation molecular
problem hazard solutions projects trace children flow temperature genes
processing site mathematics workshop ice public liquid plasma regulation
computer san nonlinear year sediment examine phase particles expression
software national spaces engineering circulation change electron magnetic function
efficient human finite faculty north management properties stars populations
programming archaeological problem mathematical flow population gas energy specific
neural october manifolds months chemical life experimental waves binding
computational earthquakes dimensional academic samples individuals temperature wave mechanisms
distributed patterns numerical equipment mantle competition spectroscopy ray evolutionary
Based on the top 20 terms in each topic, we might tentatively assign the following topic labels:
- topic 1: Computer Science
- topic 2: Archaeology
- topic 3: Mathematics
- topic 4: Academia
- topic 5: Earth Science
- topic 6: Economics
- topic 7: Chemistry
- topic 8: Physics
- topic 9: Molecular Biology
Now let's take a look at the topic-covariance matrix,
model.sigma
## Top two off-diagonal positive entries:
model.sigma[1,3] # = 18.275
model.sigma[5,9] # = 11.393
## Top two negative entries:
model.sigma[3,9] # = -27.430
model.sigma[3,5] # = -19.441According to the list above, the most closely related topics are topics 1 and 3, which correspond to the Computer Science and Mathematics topics, followed by 5 and 9, corresponding to Earth Science and Molecular Biology.
As for the most unlikely topic pairings, most strongly negatively correlated are topics 3 and 9, corresponding to Mathematics and Molecular Biology, followed by topics 3 and 5, corresponding to Mathematics and Earth Science.
The topic models so far discussed can also be used to train a classification algorithm designed to predict the topic distribution of new, unseen documents.
Let's take our 5,000 document NSF corpus, and partition it into training and test corpora,
train_corp = copy(corp)
train_corp.docs = train_corp[1:4995];
test_corp = copy(corp)
test_corp.docs = test_corp[4996:5000];Now we can train our LDA model on just the training corpus, and then use that trained model to predict the topic distributions of the five documents in our test corpus,
Random.seed!(7);
train_model = LDA(train_corp, 9)
train!(train_model, checkelbo=Inf)
test_model = predict(test_corp, train_model)The predict function works by taking in a corpus of new, unseen documents, and a trained model, and returning a new model of the same type. This new model can then be inspected directly, or using topicdist, in order to see the topic distribution for the documents in the test corpus.
Let's first take a look at both the topics for the trained model and the documents in our test corpus,
showtopics(train_model, cols=9, 20)topic 1 topic 2 topic 3 topic 4 topic 5 topic 6 topic 7 topic 8 topic 9
research system data theory research research research research plant
design research earthquake study university data project study cell
problems data project problems support project study chemistry species
systems systems research research students study data high protein
algorithms control study equations program ocean social studies cells
parallel time soil work science water understanding chemical plants
data project damage investigator award studies economic properties research
based design seismic principal scientists processes important materials studies
project analysis response project dr provide information structure genetic
models solar structures geometry sciences field policy program gene
model processing ground mathematical projects time development surface study
system information sites systems conference important work reactions molecular
analysis high analysis differential scientific climate theory electron proteins
techniques development information algebraic national marine provide metal dna
methods techniques materials groups engineering sea political experimental dr
performance computer provide space provide models science molecular genes
problem developed buildings analysis project species models systems important
computer models program methods year measurements change project understanding
work based important solutions researchers understanding scientific energy specific
developed image results finite mathematical global studies phase determine
showtitles(corp, 4996:5000) • Document 4996 Decision-Making, Modeling and Forecasting Hydrometeorologic Extremes Under Climate Change
• Document 4997 Mathematical Sciences: Representation Theory Conference, September 13-15, 1991, Eugene, Oregon
• Document 4998 Irregularity Modeling & Plasma Line Studies at High Latitudes
• Document 4999 Uses and Simulation of Randomness: Applications to Cryptography,Program Checking and Counting Problems.
• Document 5000 New Possibilities for Understanding the Role of Neuromelanin
Now let's take a look at the predicted topic distributions for these five documents,
for d in 1:5
println("Document ", 4995 + d, ": ", round.(topicdist(test_model, d), digits=3))
endDocument 4996: [0.372, 0.003, 0.0, 0.0, 0.001, 0.588, 0.035, 0.001, 0.0]
Document 4997: [0.0, 0.0, 0.0, 0.538, 0.385, 0.001, 0.047, 0.027, 0.001]
Document 4998: [0.0, 0.418, 0.0, 0.0, 0.001, 0.462, 0.0, 0.118, 0.0]
Document 4999: [0.46, 0.04, 0.002, 0.431, 0.031, 0.002, 0.015, 0.002, 0.016]
Document 5000: [0.0, 0.044, 0.0, 0.001, 0.001, 0.001, 0.0, 0.173, 0.78]
For our final model, we take a look at the collaborative topic Poisson factorization (CTPF) model.
CTPF is a collaborative filtering topic model which uses the latent thematic structure of documents to improve the quality of document recommendations beyond what would be possible using just the document-user matrix alone. This blending of thematic structure with known user prefrences not only improves recommendation accuracy, but also mitigates the cold-start problem of recommending to users never-before-seen documents. As an example, let's load the CiteULike dataset into a corpus and then randomly remove a single reader from each of the documents.
Random.seed!(1);
corp = readcorp(:citeu)
ukeys_test = Int[];
for doc in corp
index = sample(1:length(doc.readers), 1)[1]
push!(ukeys_test, doc.readers[index])
deleteat!(doc.readers, index)
deleteat!(doc.ratings, index)
endImportant. We refrain from fixing our corpus in this case, first because the CiteULike dataset is pre-packaged and thus pre-fixed, but more importantly, because removing user keys from documents and then fixing a corpus may result in a re-ordering of its user dictionary, which would in turn invalidate our test set.
After training, we will evaluate model quality by measuring our model's success at imputing the correct user back into each of the document libraries.
It's also worth noting that after removing a single reader from each document, 158 of the documents now have zero readers,
sum([isempty(doc.readers) for doc in corp]) # = 158Fortunately, since CTPF can if need be depend entirely on thematic structure when making recommendations, this poses no problem for the model.
Now that we've set up our experiment, let's instantiate and train a CTPF model on our corpus. Furthermore, in the interest of time, we'll also go ahead and GPU accelerate it.
model = gpuCTPF(corp, 100)
train!(model, iter=50, checkelbo=Inf)
## training...Finally, we evaluate the performance of our model on the test set.
ranks = Float64[];
for (d, u) in enumerate(ukeys_test)
urank = findall(model.drecs[d] .== u)[1]
nrlen = length(model.drecs[d])
push!(ranks, (nrlen - urank) / (nrlen - 1))
endThe following histogram shows the proportional ranking of each test user within the list of recommendations for their corresponding document.
Let's also take a look at the top recommendations for a particular document,
ukeys_test[1] # = 997
ranks[1] # = 0.978
showdrecs(model, 1, 120) ●●● Document 1
●●● The metabolic world of Escherichia coli is not small
...
117. #user4586
118. #user5395
119. #user531
120. #user997
What the above output tells us is that user 997's test document placed him or her in the top 2.2% (position 120) of all non-readers.
For evaluating our model's user recommendations, let's take a more holistic approach.
Since large heterogenous libraries make the qualitative assessment of recommendations difficult, let's search for a user with a small focused library,
showlibs(model, 1741) ●●● User 1741
• Region-Based Memory Management
• A Syntactic Approach to Type Soundness
• Imperative Functional Programming
• The essence of functional programming
• Representing monads
• The marriage of effects and monads
• A Taste of Linear Logic
• Monad transformers and modular interpreters
• Comprehending Monads
• Monads for functional programming
• Building interpreters by composing monads
• Typed memory management via static capabilities
• Computational Lambda-Calculus and Monads
• Why functional programming matters
• Tackling the Awkward Squad: monadic input/output, concurrency, exceptions, and foreign-language calls in Haskell
• Notions of Computation and Monads
• Recursion schemes from comonads
• There and back again: arrows for invertible programming
• Composing monads using coproducts
• An Introduction to Category Theory, Category Theory Monads, and Their Relationship to Functional Programming
The 20 articles in user 1741's library suggest that he or she is interested in programming language theory.
Now compare this with the top 25 recommendations (the top 0.15%) made by our model,
showurecs(model, 1741, 25) ●●● User 1741
1. On Understanding Types, Data Abstraction, and Polymorphism
2. Functional programming with bananas, lenses, envelopes and barbed wire
3. Can programming be liberated from the von {N}eumann style? {A} functional style and its algebra of programs
4. Monadic Parser Combinators
5. Domain specific embedded compilers
6. Type Classes with Functional Dependencies
7. Theorems for Free!
8. Scrap your boilerplate: a practical design pattern for generic programming
9. Types, abstraction and parametric polymorphism
10. Linear types can change the world!
11. Haskell's overlooked object system
12. Lazy functional state threads
13. Functional response of a generalist insect predator to one of its prey species in the field.
14. Improving literature based discovery support by genetic knowledge integration.
15. A new notation for arrows
16. Total Functional Programming
17. Monadic Parsing in Haskell
18. Types and programming languages
19. Applicative Programming with Effects
20. Triangle: {E}ngineering a {2D} {Q}uality {M}esh {G}enerator and {D}elaunay {T}riangulator
21. Motion doodles: an interface for sketching character motion
22. 'I've Got Nothing to Hide' and Other Misunderstandings of Privacy
23. Human cis natural antisense transcripts initiated by transposable elements.
24. Codata and Comonads in Haskell
25. How to make ad-hoc polymorphism less ad hoc
For the CTPF models, you may access the raw topic distributions by computing,
model.alef ./ model.betRaw scores, as well as document and user recommendations, may be accessed via,
model.scores
## M x U matrix
## M = number of documents, ordered identically to the documents in model.corp.docs.
## U = number of users, ordered identically to the keys in model.corp.users.
model.drecs
model.urecsNote, as was done by Blei et al. in their original paper, if you would like to warm start your CTPF model using the topic distributions generated by one of the other models, simply do the following prior to training your model,
ctpf_model.alef = exp.(model.beta)
## For model of type: LDA, fLDA, CTM, fCTM, gpuLDA, gpuCTM.GPU accelerating your model runs its performance bottlenecks on the GPU.
There's no reason to instantiate GPU models directly, instead you can simply instantiate the normal version of a supported model, and then use the @gpu macro to train it on the GPU,
model = LDA(readcorp(:nsf), 20)
@gpu train!(model, checkelbo=Inf)
## training...Important. Notice that we did not check the ELBO at all during training. While you may check the ELBO if you wish, it's recommended that you do so infrequently, as computing the ELBO is done entirely on the CPU.
Here are the log scaled benchmarks of the coordinate ascent algorithms for the GPU models, compared against their CPU equivalents,
As we can see, running your model on the GPU is significantly faster than running it on the CPU.
Note that it's expected that your computer will lag when training on the GPU, since you're effectively siphoning off its rendering resources to fit your model.
mutable struct Document
mutable struct Corpus
abstract type TopicModel
mutable struct LDA <: TopicModel
mutable struct fLDA <: TopicModel
mutable struct gpuLDA <: TopicModel
mutable struct CTM <: TopicModel
mutable struct fCTM <: TopicModel
mutable struct gpuCTM <: TopicModel
mutable struct CTPF <: TopicModel
mutable struct gpuCTPF <: TopicModelfunction check_doc
function check_corp
function readcorp
function writecorp
function abridge_corp!
function alphabetize_corp!
function remove_terms!
function compact_corp!
function condense_corp!
function pad_corp!
function remove_empty_docs!
function remove_redundant!
function stop_corp!
function trim_corp!
function trim_docs!
function fixcorp!
function showdocs
function showtitles
function getvocab
function getusersfunction showdocs
function showtitles
function check_model
function train!
@gpu train!
function gendoc
function gencorp
function showtopics
function showlibs
function showdrecs
function showurecs
function predict
function topicdist- Latent Dirichlet Allocation (2003); Blei, Ng, Jordan. pdf
- Filtered Latent Dirichlet Allocation: Variational Algorithm (2016); Proffitt. pdf
- Correlated Topic Models (2006); Blei, Lafferty. pdf
- Content-based Recommendations with Poisson Factorization (2014); Gopalan, Charlin, Blei. pdf
- Numerical Optimization (2006); Nocedal, Wright. Amazon
- Machine Learning: A Probabilistic Perspective (2012); Murphy. Amazon
- OpenCL in Action: How to Accelerate Graphics and Computation (2011); Scarpino. Amazon