1 Exploring Privacy Concerns in News Recommender Systems

1 Exploring Privacy Concerns in News Recommender Systems 1 Itishree 1 Department Mohallick, 1Özlem Özgöbek of Computer and Information Science Norwegian University of Science and Technology

2 Contents • • Introduction News Recommendation Characteristics of the News Domain Privacy in Recommender Systems Privacy Characteristics of the News Domain Privacy Protection Techniques in Recommender Systems Results and Discussion Conclussion

3 Have you heard of ? Introduction Information Overload

4 Introduction Dear online users, attention please !

5 Introduction Recommender Systems : A response to the problem

6 Application of Recommendation Introduction • • • Netflix Cinematch: -2/3 rented movies are due to recommendation Amazon Recommendation - 35% sales are due to recommendation provided by Amazon. Google News Personalization: - 38 % more click through due to recommendation Choicestream: - The sell of music would be more as a probable 28% user would like to buy if they found their preferred music. Facebook Friend Suggestion Linked. In Job Suggestion Celma, Ò. and Lamere, P. (2007). Music recommendation tutorial notes.

7 Introduction Dichotomy for Recommender Systems Personalized Non-Personalized Registered Customers New Customers ‘X’ visits an online store to buy a book and conducts an online product (book) search. Since ‘X’ is a registered customer, the recommendation engine gathers information related to his previous purchases (book related)- as soon as he logs in and queries for a book. The recommendation engine recommends few similar books of choice to ‘X’. Taking the context of the previous example, let us say ‘X’ is a new visitor to the online store, to make a same kind of book purchase. The recommendation engine can recommend the probable book of his choice in the form of ‘‘Best Sellers’’ despite having no information regarding ‘X’

8 Introduction Recommendation Approach Collaborative filtering (CF) technique recommends items based on user’s past behavior. o User-based CF: Finds similar users and recommends items based on their preferences. o Item-based CF: Finds similar item on the basis of user’s previous (past) preferences. Content based filtering (CBF) technique recommends items based on item features. Hybrid filtering technique combines any of the above techniques to overcome the shortcomings of any of the one technique. ADOMAVICIUS, G. and TUZHILIN, A. , 2005. Toward the Next Generation of Recommender Systems: A Survey of the State-of-the-Art and Possible Extensions. IEEE Trans. on Knowl. and Data Eng. 17, 6, 734 -749. DOI= http: //dx. doi. org/10. 1109/tkde. 2005. 99.

9 News Recommendation • The recommender systems offering news articles to online newspaper readers, based on their predicted news interest are known as news recommender systems. • Examples of News Recommender Systems include: o Google News o Yahoo! News o News. Weeder o NTNU Smart Media o PEN recsys

Characteristics of News Domain 10 What is so special about news ? • News domain is dynamic in nature. • Characristics of the news domain includes: o o o Heterogeneous nature of information sources Unstructured Format Large Volume Greater item churn Entity preferences Context Recency Filter bubble Interaction style Changing user preferences Privacy risks Explanations RICCI, F. , ROKACH, L. , SHAPIRA, B. , and KANTOR, P. B. , 2010. Recommender Systems Handbook. Springer-Verlag New York, Inc.

11 Privacy in Recommender Systems • Privacy is an important aspect of the recommender sytems as personalization is hard to achieve without any loss of privacy. • Privacy breach in recommender systems takes place due to: - Direct access to existing data - Inference of user’s preference data FRIEDMAN, A. , KNIJNENBURG, B. P. , VANHECKE, K. , MARTENS, L. , and BERKOVSKY, S. , 2015. Privacy Aspects of Recommender Systems. In Recommender Systems Handbook, F. RICCI, L. ROKACH and B. SHAPIRA Eds. Springer US, Boston, MA, 649 -688. DOI= http: //dx. doi. org/10. 1007/978 -1 -4899 -7637 -6_19.

12 Privacy in Recommender Systems Contd. . • • • Exposure risks of collected information Bias and Sabotage Shilling attack Straddlers …… RAMAKRISHNAN, N. , KELLER, B. J. , MIRZA, B. J. , GRAMA, A. Y. , and KARYPIS, G. , 2001. Privacy Risks in Recommender Systems. IEEE Internet Computing 5, 6, 54 -62. DOI= http: //dx. doi. org/10. 1109/4236. 968832

13 Privacy in Recommender Systems Cont. . Privacy risks are divided into: • Data collection • Data sales • Data retention • Employee browsing private information • Recommednation revealing information • Shared Devices or services • Stranger views private information JECKMANS, A. J. P. , BEYE, M. , ERKIN, Z. , HARTEL, P. , LAGENDIJK, R. L. , and TANG, Q. , 2013. Privacy in Recommender Systems. In Social Media Retrieval, N. RAMZAN, R. VAN ZWOL, J. -S. LEE, K. CLÜVER and X. -S. HUA Eds. Springer London, 263 -281. DOI= http: //dx. doi. org/10. 1007/978 -1 -4471 -4555 -4_12.

Privacy Characteristics of News Recommendation 14 Privacy risks in news domain • Topics mentioned in the news articles are quite diverse. • User’s preferences on news domain can reveal much more sensitive information then expected. • Broadly the privacy risks includes o The collection of user data o The management of user profiles o The generalization of personalized news paper

Privacy Characteristics of News Recommendation 15 Why should we care ? • • Reserch in this field is still young. Privacy risks in the form of o Loss of ‘personally identifiable information’ o Disclosure of page access pattern o News Context o Shared Devices o Personal news reading history accessible to strangers o News recommender systems deployed in Social Networking Sites o News service providers o Data retention

16 Privacy Protection Techniques in Recommender Systems Preserving privacy means to prevent information disclosure caused by illegitimate accces to data in the context of recommender systems. Various techniques includes: • Anonymization o k-anonymity model • Agent based approach • Perturbation (obfuscation) • Aggrgation • Differential Privacy • Cryptographic protocols • Laws and Regulations (The revision of EU data protection rules) : o Regulation EU 2016/679 o Directive EU 2016/680 o EU-US privacy shields • Awareness and user control

Results and Discussion 17 News recommendation is different. • • • Cold start Data sparsity Recency Scalabilty Serendipity Unstructured content

Results and Discussion 18 Are the aforesaid techniques applicable for news recommendation? Multiple approaches to be combined to overcome the limitations of one privacy preserving technique. • Diversification • Anonymization (recorded click histories of Google News) • Data perturbation (user can control the data for each request) • Cryptographic protocls (not recommended due to computational difficulty) • User control • Awareness • Clear and precise privacy policies should be stated by the news service providers.

19 Overall Conclusion Privacy holds a promininent place in evaluating a recommender system. Here, we have presented the state-of-the-art of the privacy cocerns and their solution in the news recommender systems with regard to recommender systems. The scope of this paper is however limited due to lack of extensive research in the said domain. A more extensive research is required to create a robust news recommender system which complies with policy, user aspects and technical aspects while considering privacy.

20 Thank You