This repository contains code for testing the possibility of personalizing trust link predictions in a multi facet framework.

1. Set $MULTIFACET_ROOT to point to the project root directory (where this file is.)
2. Follow the directions in ./settings.properties to override local settings.
3. Download the Yelp data set and filter it's contents using src/python/filter_yelp.py This has the effect of filtering out sparse users and assigning each entity a unique, consecutive integer ID. Rename the filtered files to overshadow the names of the original files and use the filtered versions going forward.
4. Use src/python/generate_single_vects.py to generate the set of single user feature vectors for users in the data set.
5. Generate pairwise comparison vectors. This can be done by running the java project with the --genPairs flag and the path to the desired output file.
6. Generate clusters of users (if desired). This can be accomplished with the src/main/python/cluster.py file and the pairwise vects generated in step 4.
7. Generate predictions for user trust links. This can be accomplished with the src/main/python/predict.py file, and the pairwise and single vect files generated in steps 3 and 4.
8. A recommender system can be trained and evaluated by calling the main java project with the path to an experiment description json file as input. See example_experiment.json for inspiration.

The experiment directory (settings.properties:experiment_dir) will be used to store intermediate files and output experiment results. experiment_dir should contain the predictions file you will use for your experiment (the one referenced in the predictionFile field of a json experiment description).

Results are placed in experiment_dir/{name}/results.txt. If multiple threads are running predictions with the same name, results are appended to this file in arbitrary order (but containing a full description of the experiment they are associated with.)