Aucun titre de diapositive

1
Broadway a recommendation computation approach
based on user behaviour similarity
B. Trousse R. Kanawati
Action AID, INRIA Sophia-Antipolis http//www.inri
a.fr/aid
2
Planning
Recommendation systems
The Broadway approach
Applications
Broadway-V1 web browsing advisor
BeCBKB query refinement advisor
3
Recommendation systems
Raw recommendations data
Recommendation computation module
Raw Recommendation data producers
Raw recommendation data collecting
Recommendation consumers
Recommendation Computing
Computed recommendations
4
Web sites recommendation approaches
Profile-Based approaches

• Content-based recommendation
• Collaboratif filtering

Data-mining based approaches

• Web-usage data analysis

Access data Users behaviour
5
The Broadway approach
Principle
Recommend to a user what others that have
behaved similarly had positively evaluated
Features
Using case based reasoning
Variable observation based behaviour modelling
6
Broadway implementation

• Modelling the user behaviour
• Determining the set of variables to observe
• Define the case structure problem and Solution.
• Define behaviours similarity measurements
• Retrieving past useful experiences
• Evaluating and adapting found solutions

Implementation CBRTools a framework for CBR
applications applied on cases with time
extended situations.
7
Broadway cycle
Session
Target case (Current behaviour, ?)
Retrieval
Target case Retrieved cases
Source case (behaviour,recommended actions)
Case base Raw observations Field knowledge
Reuse
Retain
Target case (Behaviour, adapted actions)
Target case (Behaviour, revised
recommended actions)
Recommendations
Revise
8
Applying the Broadway approach
Browsing the web Broadway-V1
Query refinement advisor Broadway-QR
Web site browsing helper.
9
Broadway-V1 user interface
10
Broadway-V1 Distributed architecture
11
Broadway-V1 case structure
Case Time-extended situation List of relevant
pages
before
Relevant page
Context
Restriction
Page address
7 8 9
5
13
3
Page content
Evaluation


Display time ratio
Time Reference
Navigation
12
Broadway-V1 experimental evaluation

• Procedure
• 2 groups with a well defined information
  searching goal
• Initialisation 20 navigations
• Results
• With Broadway Without
• Number of success 3/4 2/6
• Avg. duration 18 min 24 min
• Avg. length of navigations 19 p. 39
  p.

13
Broadway-V1 current situation future work
Study of a new version of Broadway V1 as a
multi-agent system
Validation of the prediction feature of the
Broadway approach for supporting browsing
Methodological support for the configuration of
such Broadway-based recommender systems in
specific application classes
14
Broadway-QR A QR recommender
Event server
IR server
IR client
IR server wrapper
Ex. CBKB, WSQL
QR-Recommender
Broadway-QR
In collaboration with XRCE
15
Example The BeCBKB System
16
User behaviour modelling
Solution
Context Session summery
Restriction
Query keywords
Selected Doc (Key words)
Doc. Evaluation
Query Results
Query Eval
Reference instant
17
Recommendation computation
Apply the case template on the current session at
the current instant.
1. Find past sessions with the most similar
context.
Rank solutions returned by the previous step by
using some utility function. Ex. distance form
the current query configuration
18
Broadway-QR current situation and future work
A WebSQL Wrapper is implemented
A CBKB Wrapper is under implementation
A validation study on query test data bases is
planned.
Broadway-QR as an optimiser of other query
refinement schemes
University of Torento University of Glasgow