1
Personalized Paper Recommendation Based on
User Historical Behavior
Yuan Wang
College of Information Technology Science, Nankai
University, Tianjin, China
kdd.nankai.edu.cn/wangy
Major contributors: Jie Liu, XingLiang Dong, Tianbi Liu, and YaLou Huang
Nankai University IIP Lab
Outline

Why need we provide personalized paper recommendation?
background & our motivation
 Related approaches


Personalized Paper Recommendation model
Overview
 Estimation for each part
 The optimization of the model


Results
Data Collection
 Evaluation


Conclusion
Background

the rapid development of Digital Libraries（DL）
 for
information sharing and search
 more
new idea first posted on DL
 information overload and infromation lost：
Motivation


effective technique
How to tackle the problem now？
 sending
a email
 through RSS subscription

require
user interests
explicitly
Users’ historical reveal users’ mind(*^__^*)
 browsing
 behaviors
on the site( publishing, marking a paper as
favorite, rating, making a comment, and tagging )
Motivation


Focus on providing more relevance papers to
researchers
Learn their personalized preference from their
historical behaviors
Related approaches

Personalized Recommendation
 Collaborative
 users


filtering
always prefer things their friend
PHOAKS and REFERRAL Web, CiteSeer Search Amazon，ebay ， Douban，
mainly in Commercial recommendation system
 Drawbacks

Cold Start：Pseudo Users add new score、clustering
 Content-based
 based

filtering
on content and user similarity
Web Watcher，LIRA，Leticia
Related approaches

Personalized recommendation for scientific papers
 Collaborative filtering
 Methods



citation relationship, similar to PageRank, less user preference
Learn from a recommendation system, then filtering, Input difficult
to get
get preference from log, more noise
 Drawbacks

content information should be taken consideration.
 Content-based
 take
filtering
care for messages carried with papers
 without cold start and data sparse problem
 statistical models are effective for paper recommendation
Outline

Why need we provide personalized paper recommendation?
background & our motivation
 Related approaches


Personalized Paper Recommendation model
Overview
 Estimation for each part
 The optimization of the model


Results
Data Collection
 Evaluation


Conclusion
Personalized Paper Recommendation model

A triple relationship for PPR
(Di, Dx, U)

Description
 Di
: the document set to be recommended to the user.
 Dx : the document set the user is viewing
 U : the current user

Similarity between recommended resource and users
Personalized Paper Recommendation model

Assumption
 users

and documents draw from i.i.d
Description
 p(di)
: Priori Probability of Paper
 p(dx|di) : Similarity between Papers
 p(uk|di) : Similarity between User and Paper
Priori Probability of Paper



evaluate the probability a document will be selected
through global website by users’ historical behaviors
more valuable, more operation

A = {down, keep, visit, tag, score, comment, collect, ……}

absolute discount smoothing
Similarity between Papers


title, abstract, keywords and domain of area as documents’
feature
calculate the similarity through word segmentation

w
:
tf (w; di) :
tf (di) :
tf (w, D) :
tf (D)
:

a




each word in a document
the frequency in which w appears in di
the frequency of all words in di.
the frequency that w appears in all documents
the total frequency all words appear in all documents
: a parameter used for smoothing ( 0.1 here)
Similarity between User and Paper


Actions on papers reveal users’ preference
words in users and documents draw from i.i.d

Wk

tf (w; uk) : the frequency in which w appears in profile of uk
P(w|di) : calculate as similarity between papers

: the set of each word in user’s profile
The optimization of the model

Original model did’t consider users’ preference
transitive relation
User A

User B
User C
Optimize model by Matrix Transition
The optimization of the model
normalize by colume
normalize by row
Random Walk on graph
The optimization of the model
normalize by colume
normalize by row
Random Walk on graph
Outline

Why need we provide personalized paper recommendation?
background & our motivation
 Related approaches


Personalized Paper Recommendation model
Overview
 Estimation for each part
 The optimization of the model


Results
Data Collection
 Evaluation


Conclusion
Data Collection

Train Data
 Website:
www.paper.edu.cn
 User behaviors: publish, keep, download, visit, tag,
score and make comments about papers
 Paper：first publish of papers from October 1, 2010
to March 1, 2011.

Test Data
 (U,
Dx, Di, L)
 638 data samples: 339 labeled with 1 and 299 with 0.
 26 users and 93 papers
Evaluation Matrix

MAP
 Mean
Average Precision
 show us the accuracy of models

NDCG
 Normalized
Discounted Cumulative Gain
 list accuracy evaluation based
 NDCG@1 to NDCG@6.
Results

Personalized or not
 five
percent increase
 the average improvement of NDCG is 10.2%.
Results

User preference iteration
 The
more iterations, the sharper MAP declines
 the original information is lost as the number of
iterations increases
Results

optimization of model
 “ori”
means original model, “n” means the iteration times.
 ori+1 get the highest score.
Outline

Why need we provide personalized paper recommendation?
background & our motivation
 Related approaches


Personalized Paper Recommendation model
Overview
 Estimation for each part
 The optimization of the model


Results
Data Collection
 Evaluation


Conclusion
Conclusion




Proposed a personalized recommendation model based
on users’ historical behavior.
Users’ preference profile extracted from historical
behavior, with the help of content from user model and
paper information.
Provide recommendation service with generation model
Random walk model in original model to helping
correlation transformation between users.
Appendix

MAP


M is size of recommended paper set, p (j) is the accuracy of first j recommended
papers, l (j) is label information. C (di) is the total number of related papers to
the viewed one di.
NDCG

r(i) refers to the relevant grade of ith paper and Zn is a normalized parameter,
which assures and the values NDCG@n of top results add up to 1.