Recommender Extension


The following workflows are available on myExperiment:

The following videos demonstrate the usage of the Recommender extension:

Item recommendation
Rating prediction
Attribute based prediction

Rapid Miner is data mining environment data provides tools for a diversity of data mining and data analytics tasks. Although all recommendation tasks can be cast into some more typical data mining task, state-of-the-art recommender algorithms are more apt to address typical recommendation problems (scalability, data sparsity).  Combining unique Rapid Miner environment capabilities (workflow paradigm, multitude of data processing operators for diverse data/signal types, Rapid Analytics) with state-of-the art recommender algorithms enables construction of new recommender solutions.

RecSys framework consists of two sets of operators that can be used for solving two basic types of recommendation problems – item rating prediction and item recommendation (ranking):

  • Rating prediction operators
  • Item recommendation operators

Each set of operators in RecSys has its own performance and apply model operators which are used to evaluate performance in terms of rating and ranking quality, respectively.

Rating prediction framework consists of following basic operators:

  • UserKnn: K-nearest neighbor user-based collaborative filtering using cosine-similarity (unweighted and weighted)
  • UserAttributeKnn: K-nearest neighbor user-based collaborative filtering using cosine-similarity over the user attibutes (unweighted and weighted)
  • ItemKnn: Unweighted and weighted k-nearest neighbor item-based collaborative filtering using cosine similarity
  • ItemAttributeKnn: K-nearest neighbor item-based collaborative filtering using cosine-similarity over the item attibutes
  • Random: Random item recommender for use as experimental baseline
  • MatrixFactorization: Simple matrix factorization class, learning is performed by stochastic gradient descent
  • BiasedMatrixFactorization: Matrix factorization with explicit user and item bias, learning is performed by stochastic gradient descent
  • SlopeOne: Frequency-weighted Slope-One rating prediction
  • UserItemBaseline: Baseline method for rating prediction
  • FactorWiseMatrixFactorization: Matrix factorization with factor-wise learning
  • GlobalAverage: Uses the average rating value over all ratings for prediction
  • BiPolarSlopeOne: Bi-polar frequency-weighted Slope-One rating prediction

Item recommendation framework consists of following basic operators:

  • User k-NN: K-nearest neighbor user-based collaborative filtering using cosine-similarity (unweighted and weighted)
  • User Attribute k-NN: K-nearest neighbor user-based collaborative filtering using cosine-similarity over the user attibutes (unweighted and weighted)
  • Item k-NN: Unweighted and weighted k-nearest neighbor item-based collaborative filtering using cosine similarity
  • Item Attribute k-NN: K-nearest neighbor item-based collaborative filtering using cosine-similarity over the item attibutes
  • Random: Random item recommender for use as experimental baseline
  • Biased Matrix Factorization: Matrix factorization model for item prediction (ranking) optimized for BPR
  • Weighted Regularized Matrix Factorization: Weighted matrix factorization method
  • Most Popular: Most-popular item recommender
Recommender system library MyMedia-Lite was manually ported from C# to Java RapidMiner plugin by the e-LICO consortium member Rudjer Boskovic Institute to enable workflow experiments with the state of the art recommender system algorithms for data intensive problems.
 

Recommender system extension contains operators that are suited for typical recommendation tasks:  filtering information for the particular user via :

  • Predicting most-probable ratings for the items of some DMR (rating predictions  task)
  • Ranking non-seen items by the user according to estimated (Top-N Items task)

Besides providing operators that return predictions based on user-item preference data, extension includes specific operators enabling model application and evaluation, and is accompanied with utility tools like dataset meta-data extraction operator  and workflows enabling large scale experiments for optimizing performance trade-off for specific problems.

Recommender system library MyMedia-Lite was manually ported from C# to Java RapidMiner plugin by the e-LICO consortium member Rudjer Boskovic Institute to enable workflow experiments with the state of the art recommender system algorithms for data intensive problems.
 
Source code is available on github !
 

Cite work as:

Mihelčić, M., Antulov-Fantulin, N., Bošnjak, M., Šmuc, T., Extending RapidMiner with recommender systems algorithms, RapidMiner Community Meeting and Conference 2012, Budapest, Hungary, download link

Z. Gantner, S. Rendle, C. Freudenthaler, MyMediaLite Recommender System Library, 2011.

The Recommender Extension can be downloaded from the Rapid-I Marketplace, download link

Rapid Miner Recommender Extension - user guide, download link.

Source code is available on github, download link !
 

Cite work as:

Mihelčić, M., Antulov-Fantulin, N., Bošnjak, M., Šmuc, T., Extending RapidMiner with recommender systems algorithms, RapidMiner Community Meeting and Conference 2012, Budapest, Hungary, download link

Z. Gantner, S. Rendle, C. Freudenthaler, MyMediaLite Recommender System Library, 2011.