Collaborative User Modeling
with User-generated Tags
for Social Recommender Systems
Heung-Nam Kim, et al.
Expert Systems with Applications 2011
May 25, 2011
SNU IDB Lab.
Hyunwoo Kim
Outline





Introduction
Collaborative User Modeling
Recommendation via Probabilistic Approach
Experimental Evaluation
Conclusions and Future Work
2
Introduction
 Recommender systems require information
– User’s characteristics, preferences, and needs
 Collaborative filtering
– One of the most successful technologies among recommender systems
– The best use of word-of-mouth recommendations
3
Introduction
 User model in CF
– Represented by ratings given by users on a set of items
 RS + social tagging
– A number of studies have tried to combine recommender systems with
social tagging
– Highly beneficial to both areas
– A new type of hybrid recommender systems
Content based
Hybrid
Approach
Collaborative
Filtering
4
Introduction
 Objectives
–
–
–
–
Introduce a new method of building a user model
Connect tags and ratings as a way to infer a user’s topics of interest
Provide the model with more diversity
Alleviate the cold start problem and the overspecialization issue
 Contributions
– Propose a new method of building a collaborative user model
– Present how the collaborative model can be applied to a recommender
system
– Present a new approach to identify two sets of similar neighbors by
seamlessly combining rating and tagging information
5
Collaborative User Modeling
 Relationships
– Tag-item relationships
– Informing how many times a tag has been annotated in an item
– A bag-model of social tagging indicating which tags occur and do not occur
in a particular item
 Positive and negative items
Positive items
Average rating
Negative items
6
Collaborative User Modeling
7
Collaborative User Modeling
- Identifying Tag-based Neighborhood
 Main goal
– Identify a set of users similar to a particular user
 Neighbors in terms of relevant topics are maintained to be
separated from neighbors in terms of irrelevant users
8
Collaborative User Modeling
- Enriching a User Model
 Topic
– A frequent tag pattern
web 2.0
삼성
semantic web
갤스 2
한국프로야구
ontology
슈퍼아몰레드플러스
SK 와이번스
9
박재상
Enriched topic
Collaborative User Modeling
- Enriching a User Model
10
Recommendation via Probabilistic Approach
 Item recommendation
– A classification problem either a positive class or a negative class
– Naïve Bayes classifier
positive class
negative class
Disney
3D
11
animation
Recommendation via Probabilistic Approach
12
Experimental Evaluation
 Setup
– No publicly available datasets that plentifully contain both tagging and
rating information
– Web-based interface that allows users to rate items with numerical values
and annotate them with tags
 Questions
– Is identifying two separate sets of neighbors effective for enriching user
models?
– Is the enrichment of user models effective for improving accuracy and
ranking?
– Is the quality of recommendations based on the enriched model
competitive against the existing approaches?
– Is the proposed approach able to provide proper recommendations even if
users rate few items?
13
Experimental Evaluation
 Metrics
– Precision at top N
P@10 = 7/10
answer
test
– Ranking accuracy at top N
14
Experimental Evaluation
 Neighborhood size
– Different numbers of neighbors k: 10, 20, 30, 40, 50, 60
– Accuracy of model is insensitive to the value of k
– Neighborhood with a small size provides enough to enrich topics for each
user
15
Experimental Evaluation
 Comparisons with other models
1. CF based on a user-to-user similarity (UTU)
2. CF based on an item-to-item similarity (ITI)
3. TF-IDF vector space model using tags (VSM)
16
Experimental Evaluation
 Cold-start problem
– Cold start users have less than 10 ratings
– Active users have greater than 250 ratings
– CF works well when users have abundant rating information
– # of ratings is a significant factor affecting the quality of the
recommendation in CF systems
17
Conclusions and Future Work
 What we did
– Present a method of building a user model incorporated with ratings and
tags
 What we found
– Recommendation method based on collaborative model outperforms the
initial model and baseline method
– The method produce robust performance for cold start users
 Future work
– Handling ambiguous and synonymous tags
– Supporting semantic tagging
– Social search
18