Home

Research Programme 6 (RP6)

RP6 - Personalization in Context
Programme Leaders: Prof. Barry Smyth and Prof. Alan Smeaton

Context and Objectives

The value of information resides in the ability of users and stakeholders to locate and leverage the right information at the right time, whether in pursuit of more informed mission critical decision making, or simply the enjoyment of compelling digital media. Traditional forms of information access generally place the lion’s share of effort with the end-user. For example, modern search engines expect users to locate information via carefully constructed queries and information portals expect users to browse or navigate through content categories; in both cases the end-user is responsible for initiating and guiding the hunt for relevant content.

The ability for content technologies—in particular those developed in RP4 and RP5—to harness the true potential of large-scale adaptive sensor networks will, in the end, depend critically on what might be termed the ambient findability [1] of relevant content and information. Accordingly, this research programme will develop a suite of technologies in the service of serendipitous, personalized content discovery, whereby the needs, preferences, activities and context of individuals, and groups, will be captured and harnessed to drive the proactive, high-precision recommendation of relevant content. Achieving this goal requires more than just an understanding of the content itself. The long- and short-term preferences of users, their social roles, community and group contexts, and their current and recent activities must all come to be profiled and understood as they relate to their information needs. By harnessing sensor technologies (made available through RP3) within CLARITY it will be possible, for the first time, to develop a more comprehensive picture the factors influence human preferences and intentions an so drive a new generation of personalization technologies.

Work Packages

WP 6.1: Profiling Preferences, Activities & Context

This work package will focus on the 3 core areas of profiling in CLARITY, sensing the mind, body, and place, in terms of preferences, activities, and contexts of users. In the case of the preference profiling we look at defining a wide range of ‘interest indicators’ that are suitable for online, implicit and explicit profiling, with a view to the automatic development of robust profiles that capture the preferences and needs of the individual across a variety of online domains. Moreover, sensor technologies will be used to monitor and profile the physical activities of an individual. The starting point will include the definition of new forms of ‘interest indicators’ that are suitable for activity profiling tasks within a physical setting and work will draw on the availability of a wide range of context-sensing capabilities such as location (GPS, Galileo, Egnos), activities, and proximity to others (see WP 6.3). Finally, we will also address the use of body-sensing technologies (galvanic skin response (GSR), blood pressure, heart rate, respiratory rate) with a view to profiling the physical state of individual users. Ultimately, the objective is to integrate a variety of different sensing modalities to build a comprehensive picture of a wearer’s physical state and well-being, with a view to sensing their mind, body and place states, as the basis for targeted recommendations. Such profiles will also assist the context extraction work carried out in RP4 by constraining the space of possible contexts and scenarios to be analysed.

WP 6.2: Recommendation & Collaboration

Ultimately, the multi-modal profiles learned to reflect the interests, actions and context of a user in WP 6.1 will be used as the basis for more informed and proactive decision making when it comes to the recommendation of content or actions to an individual or a group of users. In this work package we focus on the development of recommendation technologies [2, 3, 4, 5] to harness rich-profiles with a view to identifying, communicating, and explaining appropriate recommendations. This will extend conventional recommendation technologies in a number of ways. The issue of explanation in recommender systems is attracting increased levels of interest as we come to understand its importance of such annotations when it comes to helping users to appreciate the reasons behind a particular recommendation [6]. Particular attention will also be given to the often over-looked issue of when to generate a recommendation — a topic that has been largely overlooked to date on the assumption that recommendations are made at a set of points during a system interaction. We will explore a number of interrupt and messaging mechanisms as a means to provide for an unobtrusive recommendation model that blends with the user’s context; an issue of critical importance in the context of demonstrators D1 and D2.

WP 6.3: Communities, Trust & Privacy

The final work package in this research programme will addresses the relationship between individuals, groups and their profiles and the implications for recommendation and privacy. Specifically, we will explore how the profiling of individual users (their interests, actions and contexts as discussed in WP 6.1) can lead to the identification of groups and/or communities of users; that is, users who are related by context (a group) or by shared interests (a community). In particular, the ability to identify communities or groups allows personalization techniques to abstract away from the individual, in certain circumstances, in order to provide an anonymous form of personalization and recommendation that avoids many of the privacy issues that must be considered in other forms of one-to-one personalization.

Special consideration will be given to the challenge of generating recommendations for related communities and ad-hoc groups of users. The former refers to collections of users with similar interests, whereas ad-hoc groups may include users with conflicting or competing preferences. Making suggestions to a community of like-minded individuals requires an understanding of where community members share interests, whereas generating recommendations for a group of users requires a greater emphasis on the differences between member.

Finally, we will also examine the emergence ‘trust’ between community members as they participate in group/community-based recommendation scenarios and show how such models can further enhance overall recommendation performance; see for example, [7, 8]. A trust duality exists whereby trust models will be developed to support trust both between group/community members and also trust between intelligent agents which collaborate in the identification, delivery and explanation of recommendations, all to be done with ethical considerations in mind.

Novelty

The unique contribution of this work is that it extends the state-of-the-art in personalization and recommendation research a variety ways, both in the way that users are profiled and in the way that recommendations are generated and augmented. For a start, the availability of novel sensors makes it possible to profile users to a degree that has heretofore been impossible (see WP6.1); previous work in the area of profiling and personalization has been largely concerned with monitoring a limited set of online interactions only The material science and platform research of RPs 1 and 2 will provide a new range of physical sensing technologies that can be deployed via a range of platforms (e.g., wearable body-sensor networks) as per RP3, augmented by virtual sensors capable of capturing additional user context and preferences cues in RPs 4 and 5.

The novelty of WP 6.2 stems from its focus on not just the recommendations of items to users, but also the generation of explanations to justify these recommendations. Recent research has looked at annotating recommendations with information such as relevance scores (that is, the similarity of recommended items to a user’s profile in a content-based system) [2, 3, 9] or recommendation sources (that is, the related users responsible for the recommendations in a collaborative filtering system) [10, 11, 12].

The work of WP 6.2 will seek to extend such approaches by including a more explicit assessment of the trade-offs that might exist between alternative recommendations, for example. WP 6.3 will focus on personalization for groups of individuals [13, 14, 15, 16, 17], which makes particular sense in the context of pervasive information delivery services in real-world scenarios. Once again the issue of trade-offs, group dynamics, and explanation oriented recommendation splay an important role in helping group members to understand the wisdom and validity of a particular set of group recommendations.

Finally it is worth highlighting another source of novelty that arises out of the interaction between these WPs and related WPs within CLARITY. The availability of rich profiling and recommendation information can be used as a way to constrain content analysis (see RP4, for example) and information retrieval activities (RP5) or even to help modulate sensing activities within a particular sensor platform (RP3). For example, understanding that the user has engaged in a particular activity, and that she has been recommended a particular course of action, may help to constrain sensor activity and context analysis with respect to a limited set of likely scenarios.

References:

[1] P. Morville, Ambient Findability. O’Reilly Media, Inc., September 2005. [Online]. Available: http://www.amazon.co.uk/exec/obidos/ASIN/0596007655/citeulike-21
[2] M. Balabanovic and Y. Shoham, “FAB: Content-Based Collaborative Recommender.” Communications of the ACM, vol. 40(3), pp. 66–72, 1997.
[3] D. Bridge, M. Goker, L. McGinty, and B. Smyth, “Case-based recommender systems,” Knowledge Engineering Review, vol. 20, no. 3, pp. 315–320, 2006.
[4] P. Resnick and H. R. Varian, “Recommender systems,” CACM, vol. 40, no. 3, pp. 56–58, 1997.
[5] B. Smyth, K. Bradley, and R. Rafter, “Personalization techniques for online recruitment services,” Commun. ACM, vol. 45, no. 5, pp. 39–40, 2002.
[6] P. Cunningham and D. McSherry, Workshop on Explanation in CBR. Madrid, Spain: 7th European Conference on Case-Based Reasoning, 2004.
[7] J. O’Donovan and B. Smyth, “Trust in recommender systems,” in IUI ’05: Proceedings of the 10th international conference on Intelligent user interfaces. New York, NY, USA: ACM Press, 2005, pp. 167–174.
[8] ——, “Is trust robust?: an analysis of trust-based recommendation.” in Intelligent User Interfaces, 2006, pp. 101–108.
[9] B. Smyth and P. Cotter, “Surfing the DigitalWave: Generating Personalized Television Guides Using Collaborative, Case-based Recommendation,” in Proceedings of the Third International Conference on Case-based Reasoning, 1999.
[10] D. Billus and M. Pazzani, “Learning Collaborative Information Filters,” in Proceedings of the International Conference on Machine Learning, 1998, wisconsin, USA.
[11] D. Goldberg, D. Nicols, B. Oki, and D. Terry, “Using Collaborative Filtering to Weave an Information Tapestry.” Communications of the ACM, vol. 35(12), pp. 61–70, 1992.
[12] J. Konstan, B. Miller, D. Maltz, J. Herlocker, L. Gorgan, and J. Riedl, “GroupLens: Applying collaborative filtering to Usenet news,” Communications of the ACM, vol. 40(3), pp. 77–87, 1997.
[13] A. Jameson and B. Smyth, “Recommending to groups,” in The Adaptive Web: Methods and Strategies of Web Personalization, Lecture Notes in Computer Science, Vol. 4321, P. Brusilovsky, A. Kobsa, and W. Nejdl, Eds. Springer-Verlag, 2007, p. this volume.
[14] A. Jameson, S. Baldes, and T. Kleinbauer, “Two methods for enhancing mutual awareness in a group recommender system,” in Proceedings of the International Working Conference on Advanced Visual Interfaces, Gallipoli, Italy, 2004.
[15] A. Jameson, “More than the sum of its members: Challenges for group recommender systems,” in Proceedings of the International Working Conference on Advanced Visual Interfaces, Gallipoli, Italy, 2004, pp. 48–54.
[16] C. Plua and A. Jameson, “Collaborative preference elicitation in a group travel recommender system,” in Proceedings of the AH 2002 Workshop on Recommendation and Personalization in eCommerce, Malaga, Spain, 2002, pp. 148–154.
[17] J. McCarthy and T. Anagnost, “Musicfx: An arbiter of group preferences for computer aupported collaborative workouts,” in Proc. of Conference on Computer Supported Cooperative Work, 1998, pp. 363–372.