Large-scale Collaborative Ranking in Near-Linear Time

Announcement:

• The paper has been accepted for oral presentation (8.5% acceptance rate) by KDD’17, Halifax, Nova Scotia, Canada (http://www.kdd.org/kdd2017/accepted-papers).
• I gave an oral presentation about this work at Halifax sometime between August 13 - 17, 2017.
• Julia codes are available at (https://github.com/wuliwei9278/ml-1m).
• You can cite the work as

Wu, Liwei, Cho-Jui Hsieh, and James Sharpnack. "Large-scale Collaborative Ranking in Near-Linear Time." Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 2017.

Description

The code is modified from LIBPMF (https://www.cs.utexas.edu/~rofuyu/libpmf/). We extend the codebase for LIBPMF to collaborative ranking.

Platform

Unix system with g++.

Compilation

$ make

Quick start

$ ./omp-pmf-train ml1m model $ ./omp-pmf-predict ml1m/test.ratings model predicted_result

Data format

The input format is the same with LIBPMF. The input format of the training data is a directory containing a file called "meta", a file storing training ratings, and a file storing test ratings.

"meta" file contains three lines:
      1st: m n
      2nd: num_training_ratings training_file_name
      3rd: num_test_ratings test_file_name

See ml1m/ for a concrete example.

Instructions on how to run the code:

Run each program without arguments to show the detailed usage:

$ ./omp-pmf-train

Usage: omp-pmf-train [options] data_dir [model_filename]
options:
    -s type : set type of solver (default 2)
	     1 -- PirmalCR
         2 -- PrimalCR++
    -k rank : set the rank (default 10)
	-n threads : set the number of threads (default 4)
	-l lambda : set the regularization parameter lambda (default 5000)
	-t max_iter: set the number of iterations (default 10)
	-p do_predict: compute training/testing error & NDCG at each iteration or not (default 1)

$ ./omp-pmf-predict

Usage: omp-pmf-predict test_file model output_file