PeertoPeer P2P Recommender Systems RS: a research proposal

1
Peer-to-Peer (P2P)Recommender Systems (RS)a
research proposal

• Paolo Massa
• PhD student University of Trento
• paolo.massa_at_dit.unitn.it

2
Summary

• What are Recommender Systems (RS)?
• Advantages and disadvantages
• New vision (P2P)
• Which new challenges and possibilities and how to
  use them?
• Two testbeds
• A distributes scientific papers alerter
• A recommender embedded on a weblogs network

3
What is a traditional RS?

• Traditional RSs are big servers (usually web
  servers).
• The users can interact using a client (usually a
  web browser).
• The server collects information of users
  preferences on items through their interaction
  with the server, implicitly or explicitly.
• Based on these information, the server gives
  recommendations on which items the user will like.

Keep in mind Amazon.com!
4
How traditional RS recommends? (1)

• Content-based (recommend items similar to ones
  you liked)

Price
Year
Subject
Author
10
1999
Machine learning
Alice
Item1
15
2001
Phisics
Bob
Item2
56
2000
Math
Charlie
Item3
5
How traditional RS recommends? (2)

• Collaborative or Social Filtering (recommend
  items liked by users similar to you)

Item1
Item2
Item3
Item4
5
User1
3
1
5
User2
5
5

1
User3
User4
6
Problems with centralized RSs

• A RS needs to collect a lot of data from a lot of
  users.
• Users history on amazon.com cant be used on
  barnesandnobles.com (it is in their hard disks!)
• If RS server is down (temp or forever) or not
  reachable, the user cant use the service.
• Difficult to enter in the market. Big companies
  can win other markets as well (global monopoly).
• Privacy policies are defined by RS server.

7
From centralized RS to P2P RS
Every peer disposes of all the Users profiles.
8
What are Users profiles?

• A peer profile is
• Ratings on items
• (what you like)
• Ratings on peers
• (who you like/trust)
• How peers get profiles?
• Exchange
• Publish and fetch

9
Who and how generates recommendations?
IDEA Every peer can now make recommendations on
behalf of its user, using only the data it
currently has. SUB-IDEA If a peer has not
enough memory or CPU, it can ask recommendations
to other (trusted) peers.
10
Info used by RS to recommend
Every peer has the same complete information? NO

• Every peer keeps its view of the items and of
  other peers, based on
• personal interests,
• personal trust on peers,
• limitations (memory, CPU)

11
What I plan to do?

• Design new recommendation algorithms
• Facing new challenges
• The peer has to actively go fetching users
  profile from other peers and it can choose how
  frequent and from whom (information
  propagation... gossiping algorithms)
• The peer can send signed profiles of other
  peers
• Exploting new possibilities
• The peer can weight every peers contribution
  differently based on past interactions and other
  peers opinions (trust/reputation metrics).
• Identity/Privacy no more central authentication
  service, so pseudonymous based on public key

12
1st testbed A distributes scientific papers
alerter
Peer application Profile papers I like, read,
cite topics structure researchers I like
(trust) Recs Papers, Researchers, Conferences
13
1st testbed A distributes scientific papers
alerter

• STATE OF THE PROJECT
• Application defined
• Protocol coded in Java using JXTA for anonymous,
  signed and crypted P2P communication
• Graphical User Interface coded in Java
• No recommendation engine up to now (First
  release, I plan to use and extend an open source
  collaborative filtering system Alkindi)
• PROBLEM
• difficult to get widely used this application!

14
2nd testbed A recommender embedded on weblogs
network
HTTP//COCOA.ITC.IT Users can create a
compilation of classical music and download
it. USERS 1400 daily ITEMS 11.000 MP3s
(librettos, midis, ) TOPIC music (interesting
and involving) Just find a way to give direction
and tools to this army of indexers, commenters,
creators
15
2nd testbed A recommender embedded on weblogs
network

• WEBLOG (or BLOG) frequently updated, easy to
  maintain web site arranged chronologically (the
  newer entries are usually at the top).
• INTERESTING
• Google bought Blogger.com (and its company)
• Stanford and Harvard are promoting use for
  students
• Information are provided in HTML (for humans) and
  XML (for machines). Semantic web
• It is a distributed database of relationships.

16
2nd testbed A recommender embedded on weblogs
network
Peer blog Profile comments/citations on
tracks compilations, trust on other peer Recs
Tracks, Compilations, Bloggers
17
2nd testbed A recommender embedded on weblogs
network

• STATE OF PROJECT
• Setting up the blog hosting tool (Movable Type)
  and integrating in cocoa.itc.it
• Thinking ways for cocoaBloggers to provide
  semantic information.
• Thinking the recommender algorythm.

18
P2P RS

• THE END

19
What about the Items Info?

• Items Info are NOT on peers!
• They can be on a different repository
• Centralized repository (as a website)
• Distributed file storage (as Freenet or Docster)
• but even a physical library!
• Only requirement
• items are identified by a unique id (URI)

20
Ginopa communication protocol

• Peers exchange these (signed) messages
• Opinions on papers
• Ratings (peer12 rates item278 as 7/10 explicit
  on 10oct02)
• Cocitations (peer13 cites item12, item16 in my
  paper on KM)
• Codirectoring (peer45 puts item1, item6 in dir
  know manag)
• Opinions on peers (Peer31 trusts Peer12 as 0/10
  on 14Aug02)
• Recommendation requests (please, give me a
  recommendation!)
• Protocol (mainly pull to prevent DoS)
• AskMsg ReplyMsg with hash(AskMsg)
• Every msg is signed and can be safely forwarded

21
Ginopa state of project

• Defined application
• Coded in Java using JXTA protocol for anonymous,
  signed and crypted P2P communication
• Coded in Java the Graphical User Interface
• No recommendation engine up to now (First
  release, I plan to use and extend an open source
  collaborative filtering system Alkindi)

22
How traditional RS recommends?
23
How traditional RS recommends?

• Users profile is set of ratings
• Collaborative Filtering

Item1
Item2
Item1
Item1
5
User1
?
3
1
5
User2
5
5

1
User3
5
User4
24
New vision of RS exchanged info
RATINGS
(I rate Item74 as 7/10 on 23Oct02)peer31
Peer31 rates Item74 as 7/10 on 23Oct02
25
New vision of RS exchanged info
TRUSTS/REPUTATIONS
(I trust Peer12 as 0/10 on 18Aug02)peer31
Peer31 trusts Peer12 as 0/10 on 18Aug02
26
Ginopa alternative

• What if the application fails in getting widely
  used?
• Create and adapt Recommendation algorithms to the
  community of semantic web which is already
  publishing (on participants web sites) a lot of
  user profiles expressed by some semantically
  well-defined XML-based formats (RSS, RDF, FOAF,
  SMBmeta and XFML).

27
1st testbed A distributes scientific papers
alerter

• Funtionalities
• Suggest Papers
• Export HTML, XML, BibTex, RIS
• Ask Recommendations
• Express Ratings, Keywords, Summaries
• Suggest/Remind conferences (in tree)
• Suggest Users (in tree)
• Drag Drop Papers
• Manage peers trust

28
2nd testbed A recommender embedded on a weblogs
site

• USERS 1400 daily users willing to contribute
  with their passion and knowledge for music.
• ITEMS 11.000 MP3s, Composers Biographies, 400
  Opera's Librettos, 2.000 photos of composers,
  rare scores, theatres, 5.000 texts of classical
  songs in original languages, 1.100 Midi files
• TOPIC music (interesting and involving)
• Just find a way to give direction and tools to
  this army of indexers, commenters, creators.
• CocoaBloggers will
• - keep their blog
• comment on compilations,
• creating new musical itineraries (with textual
  descriptions among songs) collaboratively

29
How traditional RS recommends?

• Content-based
• (recommend items similar to ones you liked)
• Collaborative or Social Filtering
• (recommend items liked by users similar to you)