There are two operations to realize a Recommender System with Odysseus:
Example how to use the operators
In this example, the MovieLens dataset is used. The column separator was changed to tab.
The file u_ordered.data is ordered by timestamp (this is not necessary but allows implementations to that take advantage of temporal effects, e. g., concept drift).The file unique_temporal_ordered_users.data has only the user column of The file rfr.csv has a sample of the the users in u_ordered.data. Duplicates Both files are removedattached.
| Code Block | ||
|---|---|---|
| ||
#PARSER CQL#DEFINE rating_data_input_file ${PROJECTPATH}/ml-100k/u_ordered.data #DEFINE rfr_data_input_file ${PROJECTPATH}/ml-100k/rfr.csv #PARSER PQL #RUNQUERY CREATE STREAM ml100k (userid Integer, itemid Integer, rating Double, timestamp Long) WRAPPER 'GenericPull' PROTOCOL 'CSV' TRANSPORT 'File' DATAHANDLER 'Tuple' OPTIONS ( 'filename' '${PROJECTPATH}/datasets/ml-100k/u_ordered.data', 'delimiter' '\t' ,'scheduler.delay' '100' ) #RUNQUERY CREATE STREAM ml100k_users (userid Integer) WRAPPER 'GenericPull' PROTOCOL 'CSV' TRANSPORT 'File' DATAHANDLER 'Tuple' OPTIONS ( 'filename' '${PROJECTPATH}/datasets/ml-100k/unique_temporal_ordered_users.data', 'delimiter' '\t' ,'scheduler.delay' '1000' ) #PARSER PQL #ADDQUERY recommendationModels = RECOMMENDATION_LEARN( { item = 'itemid', user = 'userid', rating = 'rating', learner = 'Mahout', options = [ 'OptionRecommender'='SVDRecommender', 'OptionFactorizer'='SVDPlusPlusFactorizer' ] }, ml100k) #ADDQUERY recommendations = RECOMMENDATION( { recommender = 'recommender', user = 'userid', no_of_recommendations = 5 }, ml100k_users, recommendationModels)rating_data := ACCESS({source='rating_data', wrapper='GenericPull', transport='File', protocol='CSV', datahandler='Tuple', options=[['Delimiter', '\t'],['filename', '${rating_data_input_file}']],schema=[['user','Integer'],['item','Integer'],['rating','Double'],['timestamp','StartTimeStamp']]}) #RUNQUERY rfr := TIMEWINDOW({size = 1}, ACCESS({source='rfr', wrapper='GenericPull', transport='File', protocol='CSV', datahandler='Tuple', options=[['Delimiter', '\t'], ['filename', '${rfr_data_input_file}']], schema=[['user','Integer'], ['timestamp','StartTimeStamp']] } ) ) #QNAME RecommenderSystem #QUERY /// split learning and test data splitted_rating_data = EXTRACT_TEST_DATA({strategy = 'ITTT'}, rating_data) /// continuous learning windowed_learning_data = TIMEWINDOW({size = [30, 'days']}, 0:splitted_rating_data) models = TRAIN_RECSYS_MODEL({learner = 'BRISMF.MOA'}, windowed_learning_data) /// recommending recomm_candidates = RECOMMENDATION_CANDIDATES(JOIN(rfr, 1:models)) predicted_candidates = PREDICT_RATING(JOIN(models, recomm_candidates)) recommendations = RECOMMEND({top_n = 8, min_rating = 3.5}, predicted_candidates) /// evaluation predicted_test_data = PREDICT_RATING(JOIN(models, 1:splitted_rating_data)) model_errors = TEST_PREDICTION({aggregation_window_size = [24, 'hours']}, predicted_test_data) |