CoFiRank Release 0.1

This is CoFiRank -- Collaborative Filtering for Ranking -- code and Copyright (c) 2009 by Markus Weimer, Alexandros Karatzoglou, Alex Smola. CoFiRank includes code contributions by Quoc Le, and Choon Hui Teo.

CoFiRank is an open source library for collaborative filtering and matrix factorization licensed under the Mozilla Public License (MPL). The homepage is at:

http://www.cofirank.org

Compilation

CoFiRank requires the boost libraries to be installed on your system. By default CoFiRank also looks in the ${CoFiRank}/libs directory for boost, if a system-wide installation is not possible at your site.

CoFiRank can be compiled by either importing the code as a project into NetBeans (we used version 6.5) or at the command line using the make command:

make -f CofiRank-Makefile.mk CONF=Deploy

to compile the code without Debug information, while

make -f CofiRank-Makefile.mk CONF=Debug

includes Debug information. The binaries are then saved in dist/cofirank-deploy or dist/cofirank-debug respectively.

Running:

The code can be run on the command line as follows:

./dist/cofirank-deploy config/default.cfg

note that the first command line argument has to be a config file.

Output

Output is saved in a directory specified in the config file with the option cofi.outfolder. Output includes a file where the results of the evaluation are stored per iteration result.cvs, a file with runtime information clog.txt, a file containing the configuration options used effective-configuration.cfg, and optionally files containing the model U.lsvm, M.lsvm and the predicted output F.lsvm.

File Format for the Input Matrix

We use the format from libsvm. Each row of the input data corresponds to one row in the input matrix or one user in the cofi case. In each row col:value pairs encode the entries of this sparse matrix, where col starts with 1 and is expected to be sorted increasingly per row. Example:

1:5 39:2 65:1
12:3 26:1 77:5

This encodes two users, where user 0 has seen the movies 1,39 and 65. User 1 has seen movie 12, 26, 77

Configuration:

This is a short description of each configuration option in CoFiRank, for details on the individual configuration options please refer to the published papers.

string cofi.outfolder PATH // The folder where all output will be stored string cofibmrm.DtestFile FILENAMES // Path to the test data string cofibmrm.DtrainFile FILENAMES // Path to the train data int cofi.trainfile.size1 Positive integer // Number of rows of the train file. If not given, it will be computed. int cofi.trainfile.size2 Positive integer // Number of cols of the train file. If not given, it will be computed.

int cofi.storeU 0/1 // whether or not U, M and F shall be stored at the end int cofi.storeM 0/1 // 0 -> FALSE int cofi.storeF 0/1 // 1 -> TRUE

double cofi.minProgress 0.1 // Terminate when overall objective[t] - objective[t-1]/objective[t-1] < minProgress int cofi.minIterations 3 // Min. number of CoFi iterations over U and M int cofi.maxIterations 30 // Max number of CoFi iterations over U and M

int cofi.dimW 10 // a positive integer The number of features to learn int cofi.useAdaptiveRegularization 0/1 // Whether or not we want to use adaptive regularization double cofi.adaptiveRegularization.uExponent 1.0 // pow((#movies/max#movies),uExponent) scaling for each user-regularization value double cofi.adaptiveRegularization.wExponent 1.0 // pow((#users/max#users),wExponent) scaling for each movie-regularization value int cofi.useMovieOffset 0/1 // Enables or disables the item offset. int cofi.useUserOffset 0/1 // Enables or disables the item offset. int cofi.useGraphKernel 0/1 // whether or not top use the GraphKernel

string cofi.loss REGRESSION/ NDCG / ORDINAL // The loss to optimize for string cofibmrm.evaluation WEAK, STRONG // Evaluation in weak or strong mode double cofi.userphase.lambda 10.0 // Userphase regularization parameter lambda double cofi.moviephase.lambda 10.0 // Moviephase regularization parameter lambda

int cofi.eval.binary 0/1 // Enable disable binary classification evaluation int cofi.eval.mse 0/1 // Enable disable the RMSE evaluation int cofi.eval.ndcg 0/1 // Enable disable NDCG evaluation int cofi.eval.ndcg.k 10 // a positive integer, the truncation value in NDCG@k int cofi.eval.brmse 0/1 // Enable / disable binary rmse

double bmrm.gammaTol 0.01 // Terminate BMRM when objective[t] - objective[t-1]/objective[t-1] < gammaTol double bmrm.epsilonTol -1.0 // Terminate BMRM when objective[t] - objective[t-1] < minProgress (negative values turns this off) int bmrm.maxIter 4000 // Maximum number of BMRM iterations

int loss.ndcg.trainK 10 // Truncation value for NDCG loss double loss.ndcg.c_exponent -0.25 // c exponent for NDCG loss (see nips paper for details)