ppt slides - User Modeling Inc.

Report
22nd User Modeling, Adaptation and Personalization (UMAP 2014)
Time-Sensitive User Profile for
Optimizing Search Personalization
Ameni Kacem, Mohand Boughanem, Rim Faiz
Outline
Introduction
Related Work
Time-Based User Profile
Experiments and Results
Conclusion
Introduction
3
4
Context (1)

Personalization: search results adapted to the
user’s information needs and inetrests.
Time integration
when the words
appear
Time to discern
short- term and
long –term uer
profiles
5
Context (2)
Short-term
• Interactions extracted from the current
sesssion
• No long-term interests
Long-term
• Old interests
• No consideration of the actual user needs
6
Problem Description (1)
User interests
evolution over time
A time sensitive user profile (older
frequent
terms
should
not
outperform current and not
frequent terms).

Weight the profiles terms: both the freshness and the
frequency

Unify both the recent and persistent interests.
7
Problem Description (2)
How do temporal dynamics affect the quality
of user models in the context of personalized
search ?
How short-term (recent) profile and longterm (persistent) profile interact ?
How each of the profiles may be used in
separation or unified ?
Related Work
9
User Profiling

User profiling

Information about the user from different sources

Multiple representations:
Vector
Categories
• Weighted
keywords
• Open
Directory
Project
Semantic
network
• Concepts
10
Short-term user profile

Short-term: the interests and needs of users related to
activities of the current search session.

Daoud et al., 2009; Zemirli, 2008: the short-term user profile
is all the interactions and interests related to a single
information need.

Dumais et al., 2003; Shen et al. 1999: it represents multiple
interests emerged in a single time slot.
11
Long-term user profile

Long-term: The use of specific information: education level,
general interests, user query history and past user clickthrough
information.

Teevan et al. (2005): rich long-term user models based on
desktop search activities to improve ranking.

Tan et al. (2006): long-term language model-based
representations of users’ interests based on queries,
documents and clicks.
12
Recent Similar Works
Paper
Approach
Bennett et
al. (2012)
The first study to assess how short-term and long-term
behaviors relate, and how each may be used in isolation or
in combination.
Abel et al.
(2013)
Different strategies for mining user interest profiles from
microblogging activities ranging such as strategies that
adapt to temporal patterns that can be observed in the
microblogging behavior.
13
Temporal User Profile
Users who are not very active
• The short-term profile can eliminate relevant results
which are more related to their personal interests.
For users who are very active
• The aggregation of recent activities without ignoring
the old interests would be very interesting.
The user profile can reflect both the recurrent (persistent) and the
current (recent) interests but with different scales based on freshness.
Our Approach
15
Proposed Approach

The user profile: a vector of keywords terms corresponding to
the user interests implicitly inferred from his activities on social
Web systems.

Adjust the importance of each keyword according to the time
of its use.

Unified model:

Naturally combine short-term profile and long-term profile
into a single.

Give importance to the recent interests without ignoring the
continuous ones.
16
Main idea


Personalization in this work: weighting the user profile
keywords according to the appearing time in addition to
the frequency.
Main idea

Revising the notion of frequency by adjusting it with a
temporal function

Ensure a unified profile
17
Illustration
18
Time-Sensitive User Profile
U  W( t1( t k :)W1 ,
t 2 nTF
: W(2t k ),...,. Kt m( S : ,WS mj ) )
Sj
Sc
Sj
Sj
Sj
Sc
Sj c
Sj
19
Time-Sensitive User Profile
nTF ( t i )
Si

freq

kD
K (S , S j ) 
C
Si
freq
(ti )
Si
(t k )
Si
C


(
S
 S j )² 
1
. exp 

2 . ²
2 


Experiments and Results
21
Data Set
01-15 of december , 2013
800 Profiles
69000 Tweets
40 assessors
22
Methodology
1. Create the user profile
Sj
Sj
Sj
Sj
U  ( t1 : W1 , t 2 : W 2 ,..., t m
Sj
:Wm )

Extract the user tweets

Combine the relative frequencies with the temporal biased
function
2. Submit a query to standard search engine

Related to the user’s areas of interests defined on Twitter (800
queries)
3. Results Extraction

Top 100 Webpages
Sj
23
Methodology

Stop words removal, stemming and tokenization of
documents and users’ extracted terms (Apache Lucene
classes, Porter Stemming Filter).
3. Create the Webpage-profile WP  t wp , t wp ,..., t wp
1
2
k


Weight according to the tf-idf model
4. Rerank search results
Score ( WP , Q )   . Sim ( U , WP )  ( 1   ). Sim ( WP , Q )
24
Baselines Comparison Results
Standard Search Engine (57.87)
[email protected]
N-tf (62.68)
TSUP (78.15)
Standard Search Engine (45.67)
NDCG
N-tf (58.80)
TSUP (74.72)
Impact of User's Profile Information
Amount

25
Influence of the temporal feature: same personalization
strategy to compare the time-sensitive user profile (TSUP)
with the nTF-based user profile.
Profile
Temporal
Aspects
Short-term
Long-term
Single
Impact of User's Profile Information
Amount
26
27
Findings

Promising values:

Term frequency does not reflect the freshness of an
interest but gives an overview of how often the user
mentioned a term .

Standard search engines return relevant results to the
user query’s terms but they are indifferent to the users’
interests
Temporal function: consider the actual interests which are used
to enhance the current search without overlooking the persistent
interests and helps to personalize recurrent information needs.
28
Conclusion
29
Conclusion

Problem of personalized search: a user-modeling
framework for Twitter microblogging system.

Integration of the social data: accurate and efficient
because people are likely to write a blog or bookmark a
Webpage about something that interests them.
How the temporal-based user profile influences the
accuracy of personalized seach using a single profile
instead of separately consider the short- and long-term
user profiles?
30
Conclusion
Vector-based
representation
TemporalFrequency
Merging the term
frequency and
the freshness of
each keyword
(Kernel Function)

Encouraging results: comparison to two non-temporal sensitive
approaches.

Aggregation of the current and recurrent interests: increasing
amount of information yields to better improvement.
31
Future Work

Improve experiments: study temporal aspects when enriching
the user profile by including diverse user’s social behaviors on
the Web.

Comparison with other temporal models
Thank You For Your
Attention
AMENI KACEM
[email protected]
PHD STUDENT
PAUL SABATIER UNIVERSITY FRANCE
HIGH INSTITUTE OF MANAGEMENT,TUNISIA

similar documents