yanweifu/collranking

Preference Completion: Large-scale Collaborative Ranking from Pairwise Comparison

This repo contains the implementation of the following algorithms:

• Alternating SVM (AltSVM)
• Stochastic Gradient Descent (SGD)
• Global Ranking from All-aggregated pairwise comparisons

We use the non-convex model which is described in (3) of our paper.

Compilation

On a UNIX-based system with a C++11 supporting compiler and OpenMP API, compile using the Makefile

$ make

Experiments on numerical ratings

Our trained model can be tested in terms of NDCG@10 when the test set consists of numerical ratings.

For comparison with other rating based methods, we provide a Python script (util/num2comp.py) that divides a (user, item, rating) dataset into a training set and a test set, and extract pairwise comparisons from the training set.

1. Prepare a dataset with (user, item, ratings) triple. (Example: data/movielens1m.txt)

2. Run util/num2comp.py to get training comparisons and test ratings.

   $ python util/num2comp.py data/movielens1m.txt -o ml1m -n 50
   

   (The script also generates the training ratings which can be used for other methods)

3. Set the configuration options. (Example: config/default.cfg)

   [input]
   type = numeric
   train_file = ml1m_train.dat
   test_file = ml1m_test_ratings.lsvm
   

4. Run the binary.

   $ ./collrank
   

Experiments on binary ratings

Our trained model can also be tested in terms of Precision@K when the test set consists of binary ratings.

Please use util/bin2comp.py to divide a (user, item) dataset into a training set and a test set, and extract pairwise comparisons from the training set.

1. Prepare a dataset with (user, item) pairs. If the dataset consists of (user, item, ratings) triples, the numerical ratings are ignored. (Example: data/movielens1m.txt)

2. Run util/bin2comp.py to get training comparisons and test ratings.

   $ python util/bin2comp.py data/movielens1m.txt -o ml1m-bin -c 5000
   

3. Set the configuration options. (Example: config/default.cfg)

   [input]
   type = binary
   train_file = ml1m-bin_train.dat
   train_file_rating = ml1m-bin_train_bin.dat
   test_file = ml1m-bin_test.dat
   

   (train_file_rating is not used for traning models. It is for computing Precision@K where the training user-item pairs should be excluded)

4. Run the binary.

   $ ./collrank