This is an old revision of the document!


Problems

Known Condition:
Set {C}——users
Set {S}——possible recommended items
utility function u—— meansures the usefulness of s to user c;
matrix R ——User-Item Rating Matrix

Problem:

For each user c∈C, choose s'∈S that maximizes the user's utility.

Example:

C = {Alice, Bob, Cindy, David};
S = {K-PAX, life of Brian, Memento, Notorious};
R:
Recommander sytem predicts the unknown ratings from the matrix R.

K-PAX life of Brian Memento Notorious
Alice 4 3 2 4
Bob Φ 4 5 5
Cindy 2 2 4 Φ
David 3 Φ 5 2

Glossary

  1. Recommandation system (RS):
  2. Heuristic-based RS: define the utility function and empirically validating its performance
  3. Model-based RS: Estimating the utility function that optimizes certain performance criterion, such as the mean square error.
  4. Content-based RS: The user will be recommended items similar to the ones the user perferred in the past.
  5. Collaborative RS: The user will be recommended items that people with similar tastes and preferences liked in the past.
  6. Hybrid RS: These methods combine collaborative and content-based methods.

Papers

Tutorial

Recommendation Approach Heuristic-based Model-based
Content-based Commonly used techniques:
* TF-IDF
* Clustering
Representative research examples:
* Lang 1995
* Balabanovic 1997
* Pazzani & Billsus 1997
Commonly used techniques:
*Bayesian classifiers
* Clustering
* Decision trees
* Artificial neural networks
Representative research examples:
* Pazzan & Billsus 1997
* Mooney 1998
* Billsus& Pazzani 1999, 2000
* Zhang et al. 2002
Collaborative Commonly used techniques:
* Nearest neighbor (cosine, correlation)
* Clusstering
* Graph theory
Representative research examples
* Resnick et al. 1994
* Hill et al. 1995
* Shardanand & Maes 1995
* Breese et al. 1998
* Nakamura & Abe 1998
* Aggarwal et al. 1999
* Delgado & Ishii 1999
* Pennock & Horwitz 1999
* Sarwar et al. 2001
Commonly used techniques:
*Bayesian classifiers
* Clustering
* Decision trees
* Artificial neural networks
* Linear regression
* Problistic models
Representative research examples:
* Billsus & Pazzani 1998
* Breese et al. 1998
* Ungar & Foster 1998
* Chien & George 1999 Getoor & Sahami 1999
* Pennock & Horwitz 1999
* Goldberg et al. 2001
* Kumar et al. 2001
* Pavlov & Pennock 2002
* Shani ett al. 2002
* Yu et al. 2002, 2004
* Hofmann 2003, 2004
* Marlin 2003
* Si & Jin 2003
Hybrid Combining content-based and collaborative components by:
* Linear combination of predicted ratings
* Various voting schemes
* Incorporating one component as a part of the heuristic for the other
representative research examples:
* Balabanovic & Shoham 1997
* Claypool et al. 1999
* Good et al. 1999
* Pazzani 1999
* Billsus & Pazzani 2000
* Tran & Cohen 2000
* Melville et al. 2002
Combining content-based and collaborative components by:
* Incorporating one component as a part of the model for the other
* building one unifying model
Representative research examples:
* Basu et al. 1998
* Condliff et al. 1999
* Soboroff & Nicholas 1999
* Ansari et al. 2000
* Pospescul et al. 2001
* Schein et al. 2002

Dimensionality Reduction

    1. Abstract: Presents two different experiments to reduce the dimensionality of recommender system databases. One compares the effectiveness of system at predicting connsumer perferences, the other compares the effectiveness of sytem at producing Top-N list. Experiments shows the potential to meet the challenges of recommender systems.
    2. Explore :
    1. Abstract: Model collaborative filtering task as the classification or regression problem in machine learning and Apply SVD to reduce the dimensionality.
    2. Explore :

Article

Survey

Filter

  1. Probabilistic Matrix Factorization (PMF) (NIPS, 2008)download(pdf)—— Introduction
    1. Abstract: PMF apply a probabilistic approach using Gaussian assumptions on the knonw data and the factor matrics to factor the matrix and pridicting the missing values.Experimental resuts show that PMF perform quite well.
    2. Superiority: Scales linearly, performs well on the large, spase and imbalanced dataset.
  2. Bayesian probabilistic matrix factorization using Markov chain Monte Carlo (ICML, 2008)download(pdf)
    1. Abstract: Present a full Bayesian treatment of the PMF model in which model capacity is controlled automatically by integrating over all model parameters and hyperparameters.
    2. Superiority: higher prediction accuracy thant PMF models.
  3. Sorec: social recommendation using probabilistic matrix factorization (CIKM, 2008)download(pdf)
    1. Abstract: PMF apply a probabilistic approach using Gaussian assumptions on the knonw data and the factor matrics to factor the matrix and pridicting the missing values.Experimental resuts show that PMF perform quite well.
    2. Superiority: Scales linearly, performs well on the large, spase and imbalanced dataset.
  4. Learning to Recommend with Social Trust Ensemble (SIGIR, 2009)download(pdf)
    1. Abstract: PMF apply a probabilistic approach using Gaussian assumptions on the knonw data and the factor matrics to factor the matrix and pridicting the missing values.Experimental resuts show that PMF perform quite well.
    2. Superiority: Scales linearly, performs well on the large, spase and imbalanced dataset.

Resources

PMF:Training PMF (matlab)

Datasets

  1. MovieLens :
  2. MovieData : Sample data that contains triplets (user_id, movie_id, rating).

works

 
people/bo_xu/sorec.1291121526.txt.gz · Last modified: 2010/11/30 20:52 by box     Back to top