ssanbu08 / collabtm

collaborative topic modeling

Reference

P. Gopalan, L. Charlin, D.M. Blei, Content-based recommendations with Poisson factorization, In submission.

Installation

Required libraries: gsl, gslblas, pthread

On Linux/Unix run

./configure make; make install

Note: We have NOT tested the code on Mac. Please use Linux.

On Mac OS, the location of the required gsl, gslblas and pthread libraries may need to be specified:

./configure LDFLAGS="-L/opt/local/lib" CPPFLAGS="-I/opt/local/include" make; make install

The binary 'collabtm' will be installed in /usr/local/bin unless a different prefix is provided to configure. (See INSTALL.)

COLLABTM: Nonnegative Collaborative Topic Modeling tool

collabtm [OPTIONS]

-dir <string>            path to dataset directory with files described under INPUT below

-mdocs <int>	     number of documents

-nuser <int>	     number of users

-nvocab <int>	     size of vocabulary
	    
-k <int>                 latent dimensionality

-fixeda                  fix the document length correction factor ('a') to 1

-lda-init                use LDA based initialization (see below)

OPTIONAL:

-binary-data             treat observed ratings data as binary; if rating > 0 then rating is treated as 1

-doc-only                use document data only

-ratings-only            use ratings data only

-content-only            use both data, but predict only with the topic affinities (i.e., topic offsets are 0)

EXPERIMENTAL:

-online                  use stochastic variational inference

-seq-init -doc-only	     use sequential initialization for document only fits

RECOMMENDED

We recommend running CTPF using the following options:

~/src/collabtm -dir /mendeley -nusers 80278 -ndocs 261248 -nvocab 10000 -k 100 -lda-init -fixeda

If the document lengths are expected to vary significantly, we recommend additionally running without the "-fixeda" option above.

The above options depend on LDA-based fits being available for the document portion of the model. See below.

INPUT

We need the same files used by the basic factorization model.

train.tsv, test.tsv, validation.tsv, test_users.tsv

The new files additionally needed are mult.dat and vocab.dat. (They are really text files.) This is the "document" portion of the data. Each line of mult.dat is a document and has the following format:

 <number of words> <word-id0:count0> <word-id1:count1>....

Each line of vocab.dat is a word. Note that both the word index and the document index starts at 0. So a word-id in vocab.dat can be 0 and the document id "rated" in train.tsv can be 0.

EXAMPLE

Run two versions -- with the correction scalar 'a' inferred and one with 'a' fixed at a 1. One of these fits might be better than the other. The "-fixeda" option specifies that the documents are of similar lengths.

Always use LDA-based initialization.

~/src/collabtm -dir /mendeley -nusers 80278 -ndocs 261248 -nvocab 10000 -k 100 -lda-init

~/src/collabtm -dir /mendeley -nusers 80278 -ndocs 261248 -nvocab 10000 -k 100 -fixeda -lda-init

LDA BASED INITIALIZATION

1. Run Chong's gibbs sampler to obtain LDA fits on the word frequencies (see below for details)

2. Create a directory "lda-fits" within the "dataset directory" above and put two files in it: the topics beta-lda-k.tsv and the memberships theta-lda-k.tsv. If K=100, these files will be named beta-lda-k100.tsv and theta-lda-k100.tsv, respectively.

3. Run collabtm inference with the -lda-init option as follows (the -fixeda option fixes 'a' at 1):

~/src/collabtm -dir /mendeley -nusers 80278 -ndocs 261248 -nvocab 10000 -k 100 -lda-init

~/src/collabtm -dir /mendeley -nusers 80278 -ndocs 261248 -nvocab 10000 -k 100 -lda-init -fixeda

CHONG's GIBBS SAMPLER

The LDA code is provided under the "lda" directory.

For example, run LDA with parameters - 50 topics - the topic Dirichlet set to 0.01 - the topic proportion Dirichlet set to 0.1 as follows:

./lda --directory fit_50/ --train_data ~/arxiv/dat/mult_lda.dat --num_topics 50 --eta 0.01 --alpha 0.1 --max_iter -1 --max_time -1

mult_lda.dat contains the documents (see the David Blei's lda-c package for the exact format: http://www.cs.princeton.edu/~blei/lda-c/index.html)

Note The values of eta and alpha need to reflect those used when loading the LDA fits in CTPF (see collabtm.cc:initialize()).

The output directory ("fit_50/" in the above example) will contain the fit files which can be used to initialize CTPF with -lda-init option. Specifically *.topics corresponds to beta-lda-k.tsv, and *.doc.states corresponds to theta-lda-k.tsv.