Preparing training data

Run the mrec_prepare script to create train/test splits from a ratings dataset in TSV format. Each line should contain: user, item, score. user and item should be integer IDs starting from 1, and score is a rating or some other value describing how much the user likes or has interacted with the item. Any further fields in each line will be ignored:

$ mrec_prepare
Usage: mrec_prepare [options]

Options:
  -h, --help            show this help message and exit
  --dataset=DATASET     path to input dataset in tsv format
  --outdir=OUTDIR       directory for output files
  --num_splits=NUM_SPLITS
                        number of train/test splits to create (default: 5)
  --min_items_per_user=MIN_ITEMS_PER_USER
                        skip users with less than this number of ratings
                        (default: 10)
  --binarize            binarize ratings
  --normalize           scale training ratings to unit norm
  --rating_thresh=RATING_THRESH
                        treat ratings below this as zero (default: 0)
  --test_size=TEST_SIZE
                        target number of test items for each user, if
                        test_size >= 1 treat as an absolute number, otherwise
                        treat as a fraction of the total items (default: 0.5)
  --discard_zeros       discard zero training ratings after thresholding (not
                        recommended, incompatible with using training items to
                        guarantee that recommendations are novel)
  --sample_before_thresholding
                        choose test items before thresholding ratings (not
                        recommended, test items below threshold will then be
                        discarded)

The options are designed to support various common training and evaluation scenarios.

Rating preprocessing options

If you plan to train a SLIM recommender then you most likely need to --binarize or --normalize ratings to get good results. You may also want to set a global --rating_thresh so that an item to which a user has given a low ratings is not considered as ‘liked’ by that user; ratings below the specified threshold are set to zero.

Split options

To evaluate a recommender you need to hide some of the items that were liked by each user by removing them to a test set. Then you can generate recommendations based on the remaining training ratings, and see how many of the test items were successfully recommended. It will be hard to get meaningful results for users with very few rated items so these can simply be skipped by setting --min_items_per_user. Usually you’ll want to create several train/test splits at random using the --num_splits option and then average evaluation results across them, so that the results aren’t biased by the particular way in which any one split happens to be chosen.

You can choose how many items to move into the test set for each user with the --test_size option. A typical choice for this is 0.5, which puts half of each users liked items into the test set, but you can vary this if you need to compare with previous results that used a different split. You can also specify an absolute number of ratings by setting --test_size to an integer of 1 or more. This is also useful if you plan to measure Hit Rate in which case you should specify --test_size 1.

Filename conventions

mrec_prepare and the other mrec scripts use a set of filename conventions defined in the mrec.examples.filename_conventions module:

• input dataset: ratings.tsv
• training files created by mrec_prepare: ratings.tsv.train.0, ratings.tsv.train.1, ...
• test files created by mrec_prepare: ratings.tsv.test.0, ratings.tsv.test.1, ...
• models created by mrec_train: ratings.tsv.train.0.model.npz, ratings.tsv.train.1.model.npz, ...
• recommendations created by mrec_predict: ratings.tsv.train.0.recs.tsv, ratings.tsv.train.1.recs.tsv, ...