blei-lab / ctr Goto Github PK

ubgpu@ubgpu:~/github/ctr$ make linux
g++ -Wall -I/home/chongw/include -L/home/chongw/lib main.cpp utils.cpp corpus.cpp ctr.cpp data.cpp opt.cpp -o ctr-condor -lgsl -lm -lgslcblas
utils.cpp: In function ‘void sym_eigen(gsl_matrix_, gsl_vector_, gsl_matrix_)’:
utils.cpp:185:7: warning: variable ‘r’ set but not used [-Wunused-but-set-variable]
int r;
^
In file included from corpus.cpp:1:0:
corpus.h: In constructor ‘c_document::c_document()’:
corpus.h:20:15: error: ‘NULL’ was not declared in this scope
m_words = NULL;
^
corpus.h: In destructor ‘c_document::c_document()’:
corpus.h:34:20: error: ‘NULL’ was not declared in this scope
if (m_words != NULL) {
^
make: *_* [linux] Error 1
ubgpu@ubgpu:/github/ctr$

I cannot get the same result in the paper Collaborative Topic Modeling for Recommending Scientific articles. Please help!

I download ctr source code and lda-c source code from https://github.com/Blei-Lab, and use data from http://www.cs.cmu.edu/~chongw/data/citeulike/.

I set fixed alpha=0.005 and num_topics=200 for lda, and default parameter for ctr(no theta_opt, no lda_regression, a=1, b=0.01, lambda_u=0.01, lambda_v=100, learning_rate=-1, alpha_smooth=0, num_factors=200), and use lda's output(final.gamma, final.beta) for theta_init and beta_init. The mult.dat file I used for is from http://www.cs.cmu.edu/~chongw/data/citeulike/

Then I run ctr with cf-train-1-items.dat and cf-train-1-users.dat, and evaluate recall value with cf-test-1-users.dat, but I can only get 4.1% for recall@10 and 7.9% for recall@20.

Can you tell me what I was doing wrong? Thanks!

Is the python wrapper in the plan?

When executing ctr it dies when reading the user.dat file, it's like it cannot read the content. I have tried adding printf("%d", length); inside the while loop in data.cpp but it never enters it. Instead it just keeps on doing something until my machine runs out of memory and kills the program.

Here is my user.dat file (in txt format).
users.txt

when I look at the function c_ctr::stochastic_learn_map_estimate,I have three questions:

1. when updating the u and v, why the regulation changes from lambda_u and lambda_v to lambda_u/2n and lambda_v / (2 * items->m_vec_len[j]), which seems that it doesn't match the update equation according to stochastic Newton-Raphson rules.
2. I can't understand parameter 'lda_regression'. What does it mean?
3. In the loss function, we have $c_{ij}$. However, in the function, I cannot see the para->a and para->b in neither update rules nor the likelihood computation.
   @altosaar can you help me? Thanks.