Final (Kaggle) Score: 0.83711 Rank: 4 Kaggle name: Casper Boone

New/changed classes for this submission

• Ensemble
• EnsembleSource
• LatentFactors
• LatentFactorsCrossValidation
• LatentFactorsTestSet
• LatentFactorTraining
• Matrix
• Movie (changed)
• User (changed)
• Util (changed)

Most important features (not a complete list):

• Regularization
• Global/local biases
• Cross validation techniques
• Number of latent factors exploration, in LatentFactorTraining (tested from 1 to 29, outcome 9 or 23 are best, dependent on other parameters)
• Ensemble methods

The working of all code is explained either in JavaDoc or in in-code comments.

Final (Kaggle) Score: 0.84225 Kaggle name: Casper Boone

Most important features (not a complete list):

• Item-Item collaborative filtering
• Cosine distance (with average subtraction), also tried Jaccard and Pearson correlation. Their implementation can still be found in Predictor.
• Multi threading for improved performance (total run time is about 180s)
• Similarity caching for improved performance (slow for first iterations, but then the algorithm moves on to predicting about 40k rating per 5s)
• Global / local biases
• Light version of cross validation that gives a RMSE as a result. Which can later on also be used for training parameters. (CollaborativeFilteringTestSet)

Short overview of the multithreaded workflow: CollaborativeFiltering divides the predictions(.csv) over all predictors (all predictors are a separate thread, the number of concurrent threads is limited to 20). A Predictor is created with a given range (start to end) of predictions(.csv), for which it will predict the rating.

The working of all code is explained either in JavaDoc or in in-code comments.