Social Recommendation A Review ACM Transactions on Intelligent Systems and Technology 2013 Jiliang Tang · Xia Hu · Huan Liu Associate professor at Michigan.
2. Social Recommendation
A Review
ACM Transactions on Intelligent Systems and Technology
2013
Jiliang Tang · Xia Hu · Huan Liu
Associate professor at Michigan State University in the Computer Science and Engineering Department
3. OUTLINE
RECOMMENDATION
SYSTEM
TRADITIONAL RS Vs SRS
Existing SRS
Evaluation Metrics
Summary
POSITIVE AND
NEGATIVE
SR Classifications
INTRODUCTION
Why
Recommendation ?
01 02
04
Experience
SOCIAL
RECOMMENDER
SYSTEM
03
CONCLUSION
06
3
Future work
Research directions
05
6. 6
Recommendation
system
Collaborative
filtering (CF)
Hybrid
Memory
Based
Model
Based
Combining different
recommenders
Adding CF based characteristics
to content based models
Content Based
user-oriented
methods
item-oriented
methods
Pearson Correlation
Coefficient
Cosine similarity
Probability-based
similarity
Computing similarity for user-oriented methods
Adding content based
characteristics to CF
user-item matrix or a
sample to generate a
prediction
? 5 1
4 4 1.5
Bayesian
Bayesian belief net
Factorization
based CF
Bayesian
vector of weights
TF-IDF
Suggest items with similar keywords
8. Broad Definition
8
Narrow Definition
Social recommendation systems are systems that make recommendations
to users based on social information, such as social connections,
relationships, preferences and and behavior .
This definition encompasses systems that use only social information to
make recommendations and do not use other types of information, such as
individual preferences and behavior.
conjunction with other types of information
only social information
9. 9
Users’ preferences are likely to be similar to or influenced by their connected friends
Social relation
users are correlated when they
establish social relations
Trust influence homophily
10. 10
66% of people on social sites have asked friends or followers to help
them make a decision .
88% of links that 14-24 year olds clicked were sent to them by a friend .
78% of consumers trust peer recommendations over ads and Google
SERPs.
13. 13
Social Recommender System
Rating Matrix
Tij = 1 if uj connects to ui
users can connect to
each other
a basic CF model + a social information model
14. Existing Social Recommender Systems
14
Social Recommender
Systems
Memory based
TidalTrus
t
TrustWalker
Model based
Co-
factorizatio
n
Regulariz
ation
TidalTrust : method aims to make recommendations based not only on user preferences and
item characteristics, but also on the trustworthiness of users and items.
TidalTrust evaluates trustworthiness
TrustWalker : Uses a random walk technique to find the most trustworthy individuals in
a social network and make recommendations based on their preferences
Co-factorization: is a technique used in recommendation systems to combine multiple
matrices into a single factorized representation. The goal is to improve the accuracy of
recommendations.
Regularization: is a technique used in machine learning to prevent overfitting by adding
a penalty term to the objective function being optimized.
15. Existing Social Recommender Systems
15
Social Recommender
Systems
Memory based
TidalTrus
t
TrustWalker
Model based
Co-
factorizatio
n
Regulariz
ation
Co-factorization: is a technique used in recommendation systems to combine multiple
matrices into a single factorized representation. The goal is to improve the accuracy of
recommendations.
Regularization: is a technique used in machine learning to prevent overfitting by adding
a penalty term to the objective function being optimized.
17. Epinion03
17
This dataset provides user-item rating
information and user-user trust networks it
contains :
● 49,290 users who rated
● 139,738 different items
● 664,824 reviews and Rating
● 487,181 trust statements.
The dataset was collected from the Epinions Web site
ratings_data.txt.bz
user_id |item_id| rating_value
trust_data.txt.bz2
source_user_id |target_user_id |
trust_statement_value
22605 | 18420 | 1
https://www.shopping.com/search.html?c=Home%20%26%20Garden~~1_13000000
18. 18
Flixster
Users can rate movies and can also add some users to their friend list,establishing
social networks.
Flixster provides user-item rating information and user-user friendship networks.
CIAO
A product review website where users can rate and write reviews for various products,
and they can also establish social relations with others.
Extra contextual information
+
Epinions11
This dataset includes richer information for social recommendation, including
temporal information for both rating and social information, categories of
products, information about reviews, and distrust information, which allows
advanced research about social recommendation.
21. 21
Prediction accuracy
RMSE MAE
● Measures the closeness of predicted ratings
to the true ratings
● A smaller RMSE or MAE value means better
performance
23. Key Finding - Positive and
Negative experience
23
Can SRS improve recommendation performance ?
24. Key Finding
24
Data sparsity : The rating matrix is usually very sparse due to most users rating few of the
millions of items system is forced to choose neighbors in the small portion of comparable
users and is probably going to miss other non-comparable but relevant users
Cold star : Most of these systems are not able to generate
accurate recommendations for users with few or no ratings .
Social recommendation can make recommendations as long as
the user is connected to a large enough component of the social
network .
Recommendation can significantly reduce cold-start users.
27. 27
The Heterogeneity of Social Networks
People place trust differently to users in different domains
28. 28
● Temporal Information
● Social relations also change over time
New social relations are added, while existing social relations become inactive or
are deleted.
The changes of both ratings and social relations further exacerbate the difficulty
of exploiting temporal information for social recommendation.
Customer preferences for products drift over time !
29. 29
Cross Media Data
A user generally has multiple accounts in social media.
A new user on one website might have existed on
another website for a long time.
Integrating networks from multiple websites can bring about a huge impact on
social recommender systems and provide an efficient and effective way to solve
the cold-start problem.
We begin by giving formal definitions of social recommendation and discuss
the unique property of social recommendation and its implications compared with
those of traditional recommender systems.
A review of the existing research in the field of social recommendation
classify existing social recommender systems according to the basic models adopted to build the
systems and research directions to improve social recommendation
capabilities.
Memory based methods use either the whole user-item matrix or a sample to
generate a prediction [105], which can be further divided into user-oriented methods [44, 10] and item-oriented methods [9
Well-known model-based methods include Bayesian
belief net
Now what are the problem with customer reviews ?
not limited to online
social relations but include all kinds of available social media data such as social
tagging [69], user interactions [52] and user click behaviors
The idea is that users tend to trust and be influenced by the opinions and preferences of their friends, so incorporating information about these social connections can help improve the accuracy of recommendation algorithms. By using both the rating information (e.g. how a user rated certain items) and social information (e.g. who a user is connected to), social recommendation systems aim to capture these social connections and use them to make more personalized recommendations for users.
, traditional recommender systems assume that users are independent and
identically distributed (known as the i.i.d. assumption). However, online users are
inherently connected via various types of relations such as friendships and trust
relations
since most of SRS uses CF Following the classification of CF based
recommender systems, we classify social recommender systems into two major
categories according to their basic CF models: memory based social recommender
systems and model based social recommender systems.
since most of SRS uses CF Following the classification of CF based
recommender systems, we classify social recommender systems into two major
categories according to their basic CF models: memory based social recommender
systems and model based social recommender systems.
provide the data that is used to train and evaluate recommendation algorithms.
Epinions Epinions is a website where people can review products.
These datasets are very popular in Recommender Systems which can be used as baseline.
http://www.trustlet.org/downloaded_epinions.html
(temporal information about when ratings are provided, the category information of products, and information about the reviews such as the content and helpfulness votes
Ranking accuracy measures the effectiveness of a recommendation system by comparing the items recommended to a user and the items actually purchased by the user. Precision measures the proportion of recommended items that were purchased, while recall measures the proportion of purchased items that were recommended. F-score is a combined metric that balances precision and recall, and is less affected by the length of the recommendation list.
key
findings from both positive and negative experiences in applying social recommender systems to seek deeper understanding and further development of social
recommendation
Sparse datasets with high zero values can cause problems like over-fitting in the machine learning models and several other problems.
valuable friends, casual friends and event friends, users
are not necessarily all that similar and social relations mixed with useful and noise
connections may introduce negative information into recommender systems [92].
For example, social recommender systems simply using all available relations perform worse than traditional recommender systems
key
findings from both positive and negative experiences in applying social recommender systems to seek deeper understanding and further development of social
recommendation
u1’s social relations with {u2, u3, . . . , u9}. The user u1
may treat her social relations differently in different domains. For example, u1
may seek suggestions about “Sports” from {u2, u3}, but ask for recommendation
about “Electronics” from {u4, u5}. In [110], the authors found that people place
trust differently to users in different domains. For example, ui might trust uj in
“Sports” but not trust uj in “Electronics” at all. For different sets of items, exploiting different types of social relations can potentially benefit existing social
recommender systems [