Survey of Recommendation Systems and Algorithms

1
Survey of Recommendation Systems and Algorithms
Yuan Qu Xiaoyun Yang Tianping Huang
2
WHY?

• The amount of information available is growing
  steadily, which needs automated methods to locate
  and retrieve information with respect to users
  individual preferences.
• The number of users accessing the Internet is
  also growing, which opens up new possibilities to
  organize and recommend information.

3
OBJECTIVES

• Searching all the recommendation systems
  available from websites
• Introducing the algorithms of famous
  recommendation systems

4
INTRODUCTION

• The techniques used in todays recommendation
  systems fall into two categories
• content-based filtering
• uses actually content features of products.
• collaborative filtering
• predicts active users preference using other
  users rating, assuming that like-minded people
  tend to have similar choices.

5
Recommendation systems
6
Firefly

• Based on similarities between the interest
  profile of that user and those of other users.
• At beginning, used for music and movies, now
  extend to other media
• Characteristics
• system maintains a user profile
• word-of mouth recommendations
• vector matching based on simple rating scale

7
Tapestry

• use annotations for recommendation
• Characteristics
• first collaborative filtering
• free annotations and explicit like itor hate
  it annotations
• depends on a lot of peoples reading and voting
• hard for new areas exploration

8
GroupLens

• People who agreed in their subjective evaluation
  of past article are likely to agree again in the
  future. According to the similarity, provide
  recommendation
• Characteristics
• openness
• easy to use,vote(explicit 1-5 scale vote)
• scalability

9
Lotus Notes

• In the group, the users have similar goals and
  information interests. The system provides a
  feature to let people annotate document and send
  them to others
• Characteristics
• closed system of similar users
• annotate the documents
• use agent to represent an individual to protect
  privacy

10
Phoaks(People Helping One Another Know Stuff)

• automatically recognizing web resource
  references in a new group message and then
  attempts to classify it, then introduces it to
  other users
• Characteristics
• scan the occurrences of URLs posted in group
  messages and get the most important to users
• use implicit feedback
• role specialization
• reuse, it reuses recommendations from existing
  online conversations.

11
Pointer

• Mediator to distribute information. The pointer
  consists of URL link, contextual information, and
  optimal comments by the sender.
• Characteristics
• package contextual info with hypertext links
• easy to use( easy to add annotation)
• not anonymous

12
Mosaic

• Idea let users publish and distribute notes as
  comments added to web pages
• Characteristics
• users could publish or distribute bookmarks
• comments added to web pages

13
WebWatcher

• the server acts as tour guide for web. It
  provides interactive communication between server
  and users and provides recommendation
• Characteristics
• use previous tour, calculating the similarity of
  users
• reinforcement learning
• not the same thing as keyword-based search engine

14
GAB

• Idea collects and merges bookmark/hotlists files
  of participating users and then serves these
  files to users
• Characteristics
• ability to get users bookmarks
• multi-tree data structure
• sibling relation to avoid losing in hyperspace
• cousin relation to avoid sparse connectivity in a
  merged subject tree data base
• monitor the change of content

15
Yahoo!

• Idea manual way, one people uses tools to update
  the index as quickly as possible. Every site to
  yahoo! is examined by an expert.
• Characteristics
• expert classifiers
• user contributions, the end user also guesses and
  classifies the article

16
Fab

• Idea combination of two filters, to overcome
  problems such as, cold-start, and changing of
  users interests
• Characteristics
• update users profile, based on the content-based
  filtering
• use 7-point scale to rank
• use a series of agents to collect web pages

17
Bayesian-Mixed

• incorporates the components of both content-based
  filtering and collaborative filtering
• Characteristics
• use of all of the available data
• solve the cold-start problem

18
Mark-combination of two filters

• Idea combination of the weighted average of two
  filters( content-based and collaborative)
• Characteristics
• fully realize the advantages of two filters
• weights are determined by a per-user basis
• weights are determined by a per-item basis

19
Trend

• cold-start problem
• easy for users to participate or vote
• algorithm
• privacy

20
Algorithms on Collaborative Filtering
Example of users rating database
21
Collaborative filtering Algorithms

• Breese et al. classified CF algorithms into two
  general classes.
• Memory-based methods
• Predications are calculated over the entire
  database of users rating.
• Model-based methods
• An underlying model of users preference is first
  constructed, from which predictions are inferred.

22
Memory-based Methods

• Prediction of active users rating is a weighted
  sum of the ratings of other users.
• Weights are calculated based on

• Mean Squared Differences

• Correlation (Pearson Correlation)

• Vector Similarity

23
Improvements to Memory-based Algorithm

• Default voting
• Inverse User Frequency
• Case Amplification
• Voting by Category

24
Voting by Category
25
Model-based Methods

• Probabilistic Models
• Cluster Models
• standard cluster method Using EM algorithm to
  find clusters.
• Repeated Clustering
• Gibbs Sampling
• Bayesian Network Models
• Neural Network Models

26
Repeated Clustering
27
Other Algorithms

• A hybrid Memory- and Model-based approach
  (Personal Diagnosis)
• Assuming that all users report their ratings with
  Gaussian noise.
• Calculate the expectation of the active user
  rating.
• Have the same time-complexity as MBM.

28
Problems Improvements

• Problems
• Cold start
• High dimension, Sparse data
• Missing data
• Missing Real similarity
• Improvements
• Predictions based on users rating on latent
  features.
• Combination of CBF and CF.

29
Missing Similarity
30
Ratings to Latent Features

• Assuming that users are rating their products
  based on the latent features of products. All
  products in the database share a set of common
  features.
• Singular Value Decomposition (SVD)
• U is representative of the response of each user
  to certain features.
• V is representative of the amount of each feature
  present in each product.
• S is a matrix related to the feature importance
  in overall determination of the rating.

31
A Example of Dimension reduction using SVD
32
Recommendation Systems
Data Cleaning
Beginning
Capturing Latent Features (SVD)
Content-based Filtering
Memory-based Models
Model-based Models