Sotomayor

Document Sample
Sotomayor Powered By Docstoc
					The Design and Implementation of an
Intelligent Online Recommender System

       Authors: Rosario Sotomayor, Joe Carthy and John Dunnion




                          Speaker: Rosario Sotomayor


                  Intelligent Information Retrieval Group (IIRG)
              UCD School of Computer Science and Informatics
                           University College Dublin
                                     Ireland


 The IIRG Group                                            University College Dublin
Outlines

      An overview of Recommender Systems
      Collaborative Filtering (CF)
      Singular Value Decomposition (SVD)
      An SVD-CF Approach in the Recommender Systems Domain
      The IORS System goals
      The IORS Interface
      The IORS Architecture
      Testing Evaluation
      Conclusions/Further work



 The IIRG Group                                  University College Dublin
An overview of Recommender Systems

      What is a Recommender System?
      - Computer-based intelligent technique
      - Manages Information Overload
      - Used to efficiently provide personalized services in most e-
      commerce domains
      - Supports a customization of the customer experience through the
      representation of the products sold on a website
      - Enables the creation of a virtual world store personally designed for
      each customer
      The Goals of a Recommender System:
       - Generate suggestions about new items
       - Predict the usefulness of a specific item for a particular user

 The IIRG Group                                                    University College Dublin
An overview of Recommender Systems

      Recommender Systems in research system :

      - GroupLens
      - Movielens

      Recommender Systems in commercial use :

      - Amazon.com
      - CDNOW
      - Pandora
      - Media Unbound




 The IIRG Group                                  University College Dublin
An overview of Recommender Systems

      Amazon.com




 The IIRG Group                  University College Dublin
An overview of Recommender Systems

      Pandora:




 The IIRG Group                  University College Dublin
Outlines

      An overview of Recommender Systems
      Collaborative Filtering (CF)
      Singular Value Decomposition (SVD)
      An SVD-CF Approach in the Recommender Systems Domain
      The IORS System goals
      The IORS Interface
      The IORS Architecture
      Testing Evaluation
      Conclusions/Further work



 The IIRG Group                                  University College Dublin
Collaborative filtering (CF)

      Collaborative Filtering (CF): A promising Recommender System
      technology. Used in many of the most successful Recommender
      Systems on the web




                               r
                                          y
                                               X
                      f

                  c                   w
                           m

                                                     X

 The IIRG Group                                      University College Dublin
Collaborative filtering (CF)

      Consists of a number of Sub-Tasks:
      - Representation
      - Neighborhood formation
      - Recommendation generation
      Applications of CF:
      - E-commerce : - Amazon.com (item-to-item collaborative filtering)
                      - CDNow
      Limitations:
      - Scalability
      - Sparsity


 The IIRG Group                                          University College Dublin
Outlines

      An overview of Recommender Systems
      Collaborative Filtering (CF)
      Singular Value Decomposition (SVD)
      An SVD-CF Approach in the Recommender Systems Domain
      The IORS System Goals
      The IORS Interface
      The IORS Architecture
      Testing Evaluation
      Conclusions/Further work



 The IIRG Group                                  University College Dublin
Singular Value Decomposition (SVD)


      Singular Value Decomposition (SVD): Dimensionality reduction
      technique
            Filters the useful data from the noise in large data sets.
      Applications:
      - Information retrieval:
             Latent Semantic Indexing (LSI)
      - Recommender systems
      - Real-time signal processing
      - Seismic reflexion tomography

       Latent Semantic Indexing (LSI):

      - Synonymy: “There are many ways to refer to the same object”
      - Polysemy: “Most words have more than one distinct meaning”

 The IIRG Group                                               University College Dublin
Singular Value Decomposition (SVD)

            documents
                                                 0
                                    ·                ·       D0
                                        0
                  X          T0
   terms




                        =
                                            S0

           txd               txr            rxr               rxd

                            X = T0 · S0 · D0


 The IIRG Group                                          University College Dublin
Singular Value Decomposition (SVD)

                   X = T0 · S0 · D0
  Where:
  T0 , D0 = orthogonal matrices
   r= rank of the matrix X
   S = diagonal matrix = singular values of matrix X




 The IIRG Group                            University College Dublin
Singular Value Decomposition (SVD)


                               interesting evidence
                           0       of latent structure

                  0

                      S0
                               noise, coincidences,
                                             anomalies, …



 The IIRG Group                                          University College Dublin
Singular Value Decomposition (SVD)

            documents
                                                 0
                                    ·                ·       D0
                                        0
                  X          T0
   terms




                        =
                                            S0

           txd               txr            rxr               rxd

                            X = T0 · S0 · D0


 The IIRG Group                                          University College Dublin
Singular Value Decomposition (SVD)

           documents q
                                              0            D
                                  ·               ·
                                      0
                  X       T
   terms




                                          S

           txd            txk         kxk                kxd

                         T0·S·D0 = X  T·S·D

 The IIRG Group                                       University College Dublin
Outlines

      An overview of Recommender Systems
      Collaborative Filtering (CF)
      Singular Value Decomposition (SVD)
      An SVD-CF Approach in the Recommender Systems Domain
      The IORS System Goals
      The IORS Interface
      The IORS Architecture
      Testing Evaluation
      Conclusions/Further work



 The IIRG Group                                University College Dublin
An SVD-CF Approach in the
Recommender System Domain

      Scenario:
      - Customers and their sets of products

      Dimensionality reduction technology :
      - Singular Value Decomposition (SVD) :


      - Obtain less noisy reduced orthogonal dimensions

       - To capture latent relationships between customers and
          products
      Collaborative filtering:


     - To retrieve relevant information

 The IIRG Group                                                  University College Dublin
Outlines

      An overview of Recommender Systems
      Collaborative Filtering (CF)
      Singular Value Decomposition (SVD)
      An SVD-CF Approach in the Recommender Systems Domain
      The IORS System Goals
      The IORS Interface
      The IORS Architecture
      Testing Evaluation
      Conclusions/Further work



 The IIRG Group                                  University College Dublin
The Intelligent Online Recommender
System (IORS) goals

      Reduce the Sparsity

      Improve the quality of feedback

      Retrieval Time Reduction:
       - Timely feedback

      Search Shaping:
      - Anticipate user wishes
      - Reduce the noise generated by large quantities of data
      - Support the user in the process of selection



 The IIRG Group                                          University College Dublin
The Intelligent Online Recommender
System (IORS) goals

      Unveiling of New Preferences:
      - Customers can take advantage of new relationships among users
      and products.
      Interactive GUI feedback:
      - Filters in different fashions




 The IIRG Group                                       University College Dublin
Outlines

      An overview of Recommender Systems
      Collaborative Filtering (CF)
      Singular Value Decomposition (SVD)
      An SVD-CF Approach in the Recommender Systems Domain
      The IORS System Goals
      The IORS Interface
      The IORS Architecture
      Testing Evaluation
      Conclusions/Further work



 The IIRG Group                                  University College Dublin
The IORS Interface




                     V




                         University College Dublin
The IORS Interface




 The IIRG Group      University College Dublin
Outlines

      An overview of Recommender Systems
      Collaborative Filtering (CF)
      Singular Value Decomposition (SVD)
      An SVD-CF Approach in the Recommender Systems Domain
      The IORS System Goals
      The IORS Interface
      The IORS Architecture
      Testing Evaluation
      Conclusions/Further work



 The IIRG Group                                  University College Dublin
The IORS Architecture




 The IIRG Group         University College Dublin
Outlines

      An overview of Recommender Systems
      Collaborative Filtering (CF)
      Singular Value Decomposition (SVD)
      An SVD-CF Approach in the Recommender Systems Domain
      The IORS System Goals
      The IORS Interface
      The IORS Architecture
      Testing Evaluation
      Conclusions/Further work



 The IIRG Group                                  University College Dublin
Testing Evaluation


     Current testing is being done in order to measure the accuracy of
     SVD-CF methods. In order to do so, real data in sufficient quantity
     is being collected




The IIRG Group                                           University College Dublin
Outlines

      An overview of Recommender Systems
      Collaborative Filtering (CF)
      Singular Value Decomposition (SVD)
      An SVD-CF Approach in the Recommender Systems Domain
      The IORS System Goals
      The IORS Interface
      The IORS Architecture
      Testing Evaluation
      Conclusions/Further work



 The IIRG Group                                  University College Dublin
Conclusions/Further work

      CF is one of the most successful recommender system
      technologies, widely popular among e-tailers sites
      Recommender system technologies have become stretched by the
      huge volume of user information and are becoming even more
      stretched with the growth of Internet domain
      SVD plays a key role in the recommendation process of our system
      by addressing the gap left by collaborative filtering during the
      processing of high quantities of data
      It is important for SVD method that the derived k-dimensional factor
      space does not reconstruct the original term space perfectly, since
      the original set is deemed to be unreliable



 The IIRG Group                                           University College Dublin
Conclusions/Further work


      Further testing is required to understand the different results found
      when the k factor varies
      Further work is required to exploit SVD for item selection in order to
      find possible hidden relations among items




 The IIRG Group                                            University College Dublin
The End



                  www.cs.ucd.ie


                   Thank you!




 The IIRG Group                   University College Dublin

				
DOCUMENT INFO