Master-Seminar in Sommer Semester 2017:
Interactive Recommender Systems

(Dr. Wolfgang Wörndl, Daniel Herzog)

The goal of this seminar is to study selected aspects of recommender systems from a user’s perspective. Recommender systems recommend books, restaurants or other items to an active user or user group based on information about the users and items. The traditional focus in recommender systems research has been on the algorithms to predict ratings and generate recommendations, but has shifted more towards the user experience in recent years. Investigating the user interaction with recommender systems is especially important in mobile environments where context plays a large role.

News

[22.06.2017] The room information has been added to the schedule

[19.06.2017] The schedule for the presenations is listed below. Room information will be announced soon

[09.06.2017] Do not forget to submit your paper in the correct format until Wed, 14.06.2017, 23:59 (see information below)!

[24.04.2017] Please do not forget the information meeting on Wed, 26.04.2017, 16:00 with the seminar participants.

[27.02.2017] We assigned the topics to the students. Please check the list of topics below.

[17.02.2017] Because of the high demand on seminar places, we have increase the capacity of this seminar from 12 to 16. Topics 10-14 are new.

[02.02.2017] The new ACM templates are available. Please use the new 2017 sigconf template for your submission (see information below).

[18.01.2017] Web page online

Information

  • This seminar is for students in the Informatics Master program (module IN2107)
    • The seminar will be conducted in English language
    • Prerequisites are a Bachelor’s degree in computer science or related field
  • Students are expected to write a paper and give a presentation about the given topic area
    • Just summarizing related work is only the foundation of your paper, you need to show an own contribution beyond the summarization of other work. This contribution can be a new classification scheme, assessment and/or comparison of existing work, coming up with some novel conceptual ideas, mock-up of a new user interface, sketching new application areas and/or similar contributions
    • Your paper can be very focussed, concentrating on a (small) subset of the given area. You can for example first give an brief overview of the topic area and then dig deeper on a selected aspect
    • You need to search for suitable literature in addition to the stated references. Orient yourself to the given references or other research papers for structure/contents of your own paper. You need to cite the literature your work is based on and clearly indicate when you are adopting or paraphrasing other work
  • Information for paper
    • 7-9 pages in English in this format: http://www.acm.org/publications/proceedings-template  (2017 ACM sigconf template)
    • You can either use the LaTex (recommended) or Word templates
    • State your name, affiliation and email address (as the only author), use your own keywords and include a short abstract
    • No postal address or telephone number, no permission block, copyright line or page numbering, no categories and subject descriptors or general terms
    • References and citations need to be in the correct format (but usage of a LaTex BIB file is optional)
    • Acknowledgments or appendix are optional and not expected
  • Information for presentation
    • Duration is 25-35 minutes, plus questions&answers
    • The talk should be given freely, i.e. not completely read out from a script (in English)
    • You should present slides electronically in any format (e.g. Powerpoint or PDF)
    • You can use the TUM powerpoint template or your own format for the slides
  • The topics are either advised by Wolfgang Wörndl or Daniel Herzog. Please send an Email to your advisor for support or an appointment
  • Grading will be based on both the paper and the presentation (approximately equal weight)
  • Prerequisites for credits:
    • Submit the paper in acceptable quality until the stated deadline
    • Give a presentation of acceptable quality on the assigned date
    • Attend all presentation meetings and participate in the discussion

Procedure

  • (optional) information/pre-course meeting on Wed, 01.02.2017, 15-16h in room 01.07.023 (this meeting is completely optional and you won’t hurt your chances of getting a place if you do not come)
  • Information meeting on Wed, 26.04.2017, 16:00 with the seminar participants
  • Submission of your paper in the correct format until Wed, 14.06.2017, 23:59 (no extensions!)
    • Submit the PDF and also the source code (TeX or Word) via Email to woerndl[AT]in.tum.de, herzogd[AT]in.tum.de (sending a Dropbox link or something similar is also possible)
  • The presentations will be held on the following dates each starting at 16:00 (room to be determined): 03.07, 04.07., 05.07. and 06.07.2017

Registration

  • Registration is done using the Matching System of the department: http://www.in.tum.de/en/current-students/modules-and-courses/practical-courses-and-seminar-courses.html (you have to use this matching system to participate in the seminar!)
  • You can optionally send a short motivation statement why you want to participate in this seminar via Email to woerndl[AT]in.tum.de, herzogd[AT]in.tum.de (after 01.02.2017, max. 150 words) (sending a motivation statement is optional, but may increase your chances of getting a place)
  • You can optionally also send a list of up to 3 preferred topics via Email to woerndl[AT]in.tum.de, herzogd[AT]in.tum.de (after 01.02.2017) (sending a list of preferred topics is completely optional and we can not guarantee that you will get one of your preferences if you get a place in the seminar)

Presentation Schedule

Monday, July 3rd (room: 01.09.014)

16:00-16:45 Clemens Kamm: Explicit and Implicit User Feedback

16:45-17:30 Maria Dubinska: A Conversational and Critiquing-based Approach to Food Recommender Systems

17:30-18:15 Ali Naci Uysal: Proactive Recommendation in Mobile Guides UsingBluetooth Low Energy Devices

18:15-19:00 Marijn Jan Scholtens: A Survey on Approaches to Achieve Privacy in Location-Based Recommender Systems

Tuesday, July 4th (room: 01.11.018)

16:00-16:45 Stefan Aicher: Public Displays

16:45-17:30 Stefan Ziaras: Time-Aware Recommender Systems

17:30-18:15 Muneer Ahmad: Recommending and Presenting Sequences of Items

Wednesday, July 5th (room: 01.09.014)

16:00-16:45 Adrian Philipp: Recommendations for Groups

16:45-17:30 Vishesh Mathur: Non-Standard Context Attributes in Recommendation

17:30-18:15 Christos-Sevastianos Koliniatis: Social Recommender Systems And Their Implementations in Game Distribution Platforms

18:15-19:00 Lucia Seitz: Diversity and Serendipity in Recommender Systems

Thursday, July 6th (room: 01.09.014)

16:00-16:45 Rotem Mordoch: Machine Learning in Cross-domain Recommender Systems

16:45-17:30 Pranshu Raj Sinha: The Emotional Context: an Approach towards Increase in Choice Satisfaction and Reduction in Choice Overload in Existing Recommender Systems

17:30-18:15 Haitham Almeer Moustafa: A Usability Evaluation Strategy for Recommender Systems

Topics

1. Collecting Implicit and Explicit Feedback for Recommendations (Presenter: Clemens Kamm, Advisor: Daniel Herzog)

Many recommendation techniques such as collaborative filtering need user ratings to recommend items. User ratings can be created by collecting explicit and explicit feedback. Explicit feedback, e.g., a user rating on a scale from 1 to 5 is very accurate but it demands an effort from the user. On the other hand, collecting implicit feedback is a bigger challenge but observing the user’s browsing behavior or eye movements when interacting with a recommender system allows to identify recommendations that better satisfy the user’s needs without annoying the user.

  • Schafer et al. (2007): Collaborative Filtering Recommender Systems (Section 9.4)
  • Jawaheer (2010): Comparison of Implicit and Explicit Feedback from an Online Music Recommendation Service
  • Sparling and Sen (2011): Rating: How Difficult is It?

2. Conversational and Critique-Based Recommender System (Presenter: Maria Dubinska, Advisor: Daniel Herzog)

Structured dialogues between the user and the recommender system promise better recommendations than traditional recommendations based on a one-time request/response. Recommender system providing dialogues to iteratively learn the user’s preferences and improve the recommendations are called conversational. A conversational recommender system can either suggest two or more alternatives for concrete recommendations and let the user indicate her or his preferences for one item over the others or iteratively ask the user questions about the most important features of the expected recommendations.

  • Shimazu (2001): ExpertClerk: navigating shoppers' buying process with the combination of asking and proposing
  • Mahmood and Ricci (2009): Improving Recommender Systems with Adaptive Conversational Strategies
  • Chen an Pu (2011): Critiquing-based recommenders: survey and emerging trends 

3. Proactive Recommendation in Mobile Guides (Presenter: Ali Naci Uysal, Advisor: Wolfgang Wörndl)

Proactivity means that the systems pushes suitable items to the user, without explicit user request. So the question is not only which item to recommend, but also when. The decision can be made based on the current context, e.g. time and location. For example, a user visiting a city gets recommendations for nearby recommended restaurants around lunch time. The question is also when and how to communicate the items to the user. In addition, explanations of recommendations are important when recommendations are proactively delivered because the user may be unaware why she received recommendations.

  • Wörndl at al. (2011): A Model for Proactivity in Mobile, Context-aware Recommender Systems
  • Braunhofer et al. (2015): A Context-aware Model for Proactive Recommender Systems in the Tourism Domain
  • Sabic and Zanker (2015): Investigating User's Information Needs and Attitudes Towards Proactivity in Mobile Tourist Guides

4. Privacy-Enhanced Recommender Systems  (Presenter: Marijn Jan Scholtens, Advisor: Wolfgang Wörndl)

Recommender systems generate personalized recommendation based on information about users. The more accurate this information is, the better recommendations can be tailored towards user needs and interests. But collecting and utilizing personal data raises privacy issues. Users may be unaware which data is collected and do not want systems to acquire information about themselves. There are existing solutions to generate personalized recommendation while still respecting user privacy. Thus, this is topic about the trade-off between personalization and privacy in recommender systems.

  • Drosatos et al. (2015): Pythia: A Privacy-Enhanced Personalized Contextual Suggestion System for Tourism
  • Saravanan and Ramakrishnan (2016): Preserving Privacy in the Context of Location Based Services Through Location Hider in Mobile-Tourism
  • Friedman et al. (2016): A Differential Privacy Framework for Matrix Factorization Recommender Systems

5. Distributed User Interfaces (Presenter: Tobias Zappe, Advisor: Daniel Herzog)

Distributed user interfaces allow computer interfaces to be distributed across multiple devices, multiple users, and multiple platforms. Distributed user interfaces are a current topic in the research of computer science and human computer interaction but are already well-established in different scenarios. One examples is a media player than runs on a TV but can be controlled by a smartphone which is connected to the TV. This scenario is also a good example for a multi-user application enabled by distributed user interfaces: Multiple users could, for example, share their movie preferences on their personal devices and the software running on the TV could choose a movie suitable for all users.

  • Vanderdonckt (2010): Distributed User Interfaces: How to Distribute User Interface Elements across Users, Platforms, and Environments
  • Elmqvist (2011): Distributed User Interfaces: State of the Art
  • Abdrabo and Wörndl (2016): DiRec: A Distributed User Interface Video Recommender 

6. Public Displays (Presenter: Stefan Aicher, Advisor: Daniel Herzog)

Public displays are large displays in public spaces that allow a community to interact with the screen and / or other users. Public displays are getting more and more popular as they can provide useful information such as maps, information about public transport or interesting spots nearby to a person passing by or a crowd. Recent work tries to improve the services proved by public screens by enhancing the means of interaction and enabling personalized content on the displays. Nevertheless, these innovations make some people hesitate interacting with such screens because of social embarrassment or privacy issues.

  • Brignull and Rogers (2003): Enticing People to Interact with Large Public Displays in Public Spaces
  • Boring et al. (2009): Scroll, Tilt or Move It: Using Mobile Phones to Continuously Control Pointers on Large Public Displays
  • Alt et al. (2012): How to Evaluate Public Displays 

7. Time-Aware Recommender Systems (Presenter: Stefan Ziaras, Advisor: Wolfgang Wörndl)

Context such as time and location is important information to be considered in recommender systems. A simple example is a recommendation service for nearby restaurants based on the current user location. But time also plays a larger role. For example, there are temporal constraints when visiting a city because some sights may have opening times and a restaurant for dinner does not make much sense in the morning. In addition, user preference may be dynamic and change over time.

  • Campos et al. (2014): Time-aware Recommender Systems: a Comprehensive Survey and Analysis of Existing Evaluation Protocols
  • Aggrawal (2016): Time- and Location-Sensitive Recommender Systems
  • Yuan et al. (2013): Time-aware Point-of-interest Recommendation

8. Mobile Recommender Systems (Presenter: Ahmet Melih Bercin, Advisor: Wolfgang Wörndl)

Mobile devices such as smartphones are increasingly used for information access tasks while traveling. However, mobile information access still suffers from limited resources regarding input capabilities, displays, network bandwidth and other limitations of mobile devices. In addition, mobile applications must consider mobile user constraints such as limited attention span while moving, changing locations and contexts, and expectations of quick and easy interactions. Therefore, it is desirable to tailor information access to the current user needs in mobile recommendation and other adaptive systems.

  • Lathia (2015): The Anatomy of Mobile Location-Based Recommender Systems
  • Gavalas et al. (2014): Mobile Recommender Systems in Tourism
  • Baltrunas at al. (2012): Context Relevance Assessment and Exploitation in Mobile Recommender Systems

9. Recommending and Presenting Sequences of Items (Presenter: Muneer Ahmad, Advisor: Daniel Herzog)

Many recommender systems recommend single items such as movies or restaurants. Recommending a sequence of items, for example a music playlist or a tourist trip composed of multiple points of interest, is a more complicated issue. Not only the choice of items but also the sequence order influences the quality of the recommendation. For example, a strong ending with a very well-liked item at the end of a sequence might maximize the user satisfaction as the user tends to remember the end of recommendation most. When a recommendation is found, the system has to decide how to present the sequence to the user. The whole sequence could be recommended at a time but in certain scenarios, it might be appreciated if only the upcoming item is presented to the user at the right time.

  • Masthoff (2004): Group Modeling: Selecting a Sequence of Television Items to Suit a Group of Viewers
  • Masthoff (2015): Group Recommender Systems: Aggregation, Satisfaction and Group Attributes
  • Wörndl et al. (2017): Recommending a Sequence of Interesting Places for Tourist Trips 

10. Recommendations for Groups (Presenter: Adrian Philipp, Advisor: Daniel Herzog)

When recommending items to a group of users instead of a single user, the preferences of all group members have to be taken into account. Different preference aggregation strategies exist for this purpose. Basic approaches such as calculating an average preference of all users are easy to realize but might not be optimal. For example, if one user really dislikes a certain item, this item should not be recommended even if the majority of group members likes the item. It is important to say that there is no perfect way to aggregate the individual preferences. Instead, the group’s intrinsic characteristics and the problem’s nature have to be considered.

  • Yu et al. (2006): TV Program Recommendation for Multiple Viewers Based on User Profile Merging
  • Jameons and Smyth (2007): Recommendation to Groups
  • Masthoff (2015): Group Recommender Systems: Aggregation, Satisfaction and Group Attributes 

11. Non-Standard Context Attributes in Recommendation (Presenter: Vishesh Mathur, Advisor: Wolfgang Wörndl)

The goal of context-aware recommender systems is to tailor recommendations towards the current situation of users (also see topic 7). However, the considered context is often limited to location and time. The goal of this topic is to discuss the integration of non-standard context attributes in recommendation approaches, i.e. context other location and time. These may be inferred from sensor data and may include user emotions and mood, current activity and health-related information, traffic and transportation data, learning contexts, and others. 

  • Zheng et al. (2013): The Role of Emotions in Context-aware Recommendation
  • Ilarri et al. (2015): A Review of the Role of Sensors in Mobile Context-Aware Recommendation Systems
  • Verbert et al. (2012): Context-Aware Recommender Systems for Learning: A Survey and Future Challenges

12. Social Recommender Systems (Presenter: Christos-Sevastianos Koliniatis, Advisor: Wolfgang Wörndl)

Social recommender systems make use of the huge content in the social media domain which provide platforms for users to create, annotate and share information. The information in the social network can then be used to improve the recommendation process itself and provide a more personalized and social experience. In addition, social recommendation also aims at suggesting attractive and relevant social media items such as blog posts, images shared by other users, or even other people, e.g. experts for a certain task. 

  • Aggrawal (2016): Social and Trust-Centric Recommender Systems
  • Guy (2015): Social Recommender Systems
  • Guy et al. (2010): Social Media Recommendation Based on People and Tags 

13. Diversity and Serendipity in Recommender Systems (Presenter: Lucia Seitz, Advisor: Daniel Herzog)

Many recommender systems try to find very accurate recommendations by suggesting items which are very similar to the user query. Often users expect a set of recommendations from which they can choose their favorite recommendation. If all recommendations in this set are very similar to each other, a lot of attractive items could be overlooked. This is why state-of-the-art recommender should not only recommend similar items. Instead, diversity and serendipity are important metrics to ensure that recommendations are not only accurate, but also attractive and surprising to the users. 

  • Smyth (2007): Case-Based Recommendation (Section 11.3)
  • Ge et al. (2010): Beyond Accuracy: Evaluating Recommender Systems by Coverage and Serendipity
  • Vargas and Castells (2011): Rank and Relevance in Novelty and Diversity Metrics for Recommender Systems

14. Cross-domain Recommendations (Presenter: Rotem Mordoch, Advisor: Daniel Herzog)

The majority of recommender systems offer recommendations for items belonging to a single domain such as movies, restaurants or books. For large e-commerce sites like Amazon, recommending products of different domains on the basis of one single user profile is very attractive as it offers an incentive for customers to purchase multiple products at the same time. Different techniques for aggregating knowledge from different domains, linking and transferring knowledge between domains and to use it for cross-domain recommendations exist. Even though cross-domain recommendations are an emerging topic in research, it remains very challenging to realize.

  • Winoto and Tang (2008): If You Like the Devil Wears Prada the Book, Will You also Enjoy the Devil Wears Prada the Movie? A Study of Cross-Domain Recommendations
  • Fernández-Tobías et al. (2012): Cross-domain recommender systems: A survey of the state of the art
  • Cantador et al. (2015): Cross-Domain Recommender Systems

15. Human Decision Making and Recommender Systems (Presenter: Pranshu Raj Sinha, Advisor: Wolfgang Wörndl)

The ultimate goal of recommender systems is to support users in decision making. Systems try to predict ratings for products or present a ranked list of items. However, the item with the highest predicted rating may not be the one a user choses as other aspects have to be considered. For example, how items are presented, the number of items in a list etc. have been shown to influence users' decision making.

  • Jameson et al. (2015): Human Decision Making and Recommender Systems
  • Bollen at al. (2010: Understanding Choice Overload in Recommender Systems
  • Ekstrand and Willemsen (2016): Behaviorism is Not Enough: Better Recommendations Through Listening to Users

16. User-Centric Evaluation (Presenter: Haitham Almeer Moustafa, Advisor: Wolfgang Wörndl)

The main options of recommender systems evaluation are online and offline studies. Offline experiments work on existing data sets but do not take into account how the recommendations affect user behavior. In online studies, users can test a system in a real setting so that more realistic experiences can be gathered and analyzed. There is a lot of work on methods for user-centric evaluation in general and some frameworks deal more specifically with how to evaluate recommender systems in online, user-centric studies.

  • Knijnenburg and Willemsen (2016): Evaluating Recommender Systems with User Experiments
  • Dooms at al. (2011): A User-Centric Evaluation of Recommender Algorithms for an Event Recommendation System
  • Pu and Chen (2010): A User-Centric Evaluation Framework of Recommender Systems

Contact