MyMediaLite Item Recommendation from Positive-Only Feedback 3.11
usage: item_recommendation --training-file=FILE --recommender=METHOD [OPTIONS]
methods (plus arguments and their defaults):
- BPRMF num_factors=10 bias_reg=0 reg_u=0.0025 reg_i=0.0025 reg_j=0.00025 num_iter=30 learn_rate=0.05 uniform_user_sampling=True with_replacement=False update_j=True
supports --find-iter=N, --online-evaluation
- ItemAttributeKNN k=80 correlation=Cosine q=1 weighted=False alpha=0.5 (only for BidirectionalConditionalProbability)
needs --item-attributes=FILE
supports --online-evaluation
- ItemKNN k=80 correlation=Cosine q=1 weighted=False alpha=0.5 (only for BidirectionalConditionalProbability)
supports --online-evaluation
- MostPopular by_user=False
supports --online-evaluation
- Random
- UserAttributeKNN k=80 correlation=Cosine q=1 weighted=False alpha=0.5 (only for BidirectionalConditionalProbability)
needs --user-attributes=FILE
supports --online-evaluation
- UserKNN k=80 correlation=Cosine q=1 weighted=False alpha=0.5 (only for BidirectionalConditionalProbability)
supports --online-evaluation
- WRMF num_factors=10 regularization=0.015 alpha=1 num_iter=15
supports --find-iter=N, --online-evaluation
- Zero
- MultiCoreBPRMF num_factors=10 bias_reg=0 reg_u=0.0025 reg_i=0.0025 reg_j=0.00025 num_iter=30 learn_rate=0.05 uniform_user_sampling=False with_replacement=False update_j=True max_threads=100
supports --find-iter=N, --online-evaluation
- SoftMarginRankingMF num_factors=10 bias_reg=0 reg_u=0.0025 reg_i=0.0025 reg_j=0.00025 num_iter=30 learn_rate=0.1 uniform_user_sampling=True with_replacement=False update_j=True
supports --find-iter=N, --online-evaluation
- WeightedBPRMF num_factors=10 bias_reg=0 reg_u=0.0025 reg_i=0.0025 reg_j=0.00025 num_iter=30 learn_rate=0.05
supports --find-iter=N, --online-evaluation
- MostPopularByAttributes
needs --item-attributes=FILE
- BPRSLIM reg_i=0.0025 reg_j=0.00025 num_iter=15 learn_rate=0.05 uniform_user_sampling=True with_replacement=False update_j=True
supports --find-iter=N, --online-evaluation
- LeastSquareSLIM reg_l1=0.01 reg_l2=0.001 num_iter=15 K=50
supports --find-iter=N, --online-evaluation
- ExternalItemRecommender prediction_file=FILENAME
method ARGUMENTS have the form name=value
general OPTIONS:
--recommender=METHOD use METHOD for recommendations (default: MostPopular)
--recommender-options=OPTIONS use OPTIONS as recommender options
--help display this usage information and exit
--version display version information and exit
--random-seed=N initialize random number generator with N
files:
--training-file=FILE read training data from FILE
--test-file=FILE read test data from FILE
--file-format=ignore_first_line|default
--no-id-mapping do not map user and item IDs to internal IDs, keep the original IDs
--data-dir=DIR load all files from DIR
--user-attributes=FILE file with user attribute information, 1 tuple per line
--item-attributes=FILE file with item attribute information, 1 tuple per line
--user-relations=FILE file with user relation information, 1 tuple per line
--item-relations=FILE file with item relation information, 1 tuple per line
--save-model=FILE save computed model to FILE
--load-model=FILE load model from FILE
--save-user-mapping=FILE save user ID mapping to FILE
--save-item-mapping=FILE save item ID mapping to FILE
--load-user-mapping=FILE load user ID mapping from FILE
--load-item-mapping=FILE load item ID mapping from FILE
data interpretation:
--user-prediction transpose the user-item matrix and perform user prediction instead of item prediction
--rating-threshold=NUM (for rating data) interpret rating >= NUM as positive feedback
choosing the items for evaluation/prediction (mutually exclusive):
--candidate-items=FILE use items in FILE (one per line) as candidate items
--overlap-items use only items that are both in the training and the test set as candidate items
--in-training-items use only items in the training set as candidate items
--in-test-items use only items in the test set as candidate items
--all-items use all known items as candidate items
The default is to use both the items in the training and the test set as candidate items.
choosing the users for evaluation/prediction
--test-users=FILE predict items for users specified in FILE (one user per line)
prediction and evaluation:
--predict-items-number=N predict N items per user
--repeated-items items accessed by a user before may be in the recommendations (and are not ignored in the evaluation)
prediction:
--prediction-file=FILE write ranked predictions to FILE, one user per line
evaluation:
--cross-validation=K perform k-fold cross-validation on the training data
--test-ratio=NUM evaluate by splitting of a NUM part of the feedback
--num-test-users=N evaluate on only N randomly picked users (to save time)
--online-evaluation perform online evaluation (use every tested user-item combination for incremental training)
--compute-fit display fit on training data
--measures=LIST the evaluation measures to display (default is 'AUC, prec@5')
use --help-measures to get a list of all available measures
finding the right number of iterations (iterative methods)
--find-iter=N give out statistics every N iterations
--num-iter=N start measuring at N iterations
--max-iter=N perform at most N iterations
--epsilon=NUM abort iterations if main measure is less than best result plus NUM
--cutoff=NUM abort if main measure is below NUM
