Collaborative Technologies & Social Media Lab Binzmühlestr. 14, CH- 8050 Zurich
[email protected] +41-44-635-7137
ABSTRACT
This workshop contribution presents experiences from our work on supporting exploratory search tasks by using visualization to augment the search process and information sensemaking in a specific domain of cross-community knowledge exchange. We then discuss how the lessons learned here are related and can be transferred to supporting exploratory access to user-generated media collections. The first part presents three main results of our recent work: 1) a concrete visual information interface for explorative, multi-perspective access to community information spaces, 2) an evaluation framework for measuring search performance in terms of user-centered knowledge construction effects and 3) the results of an empirical laboratory study applying this framework to compare the use of a visual information interface (Knowledge Explorer) and a traditional search system (Google). The second part discusses how these experiences are linked to the currently exploding phenomenon of personal and participatory media where huge amounts of information are increasingly produced and shared between users, in a number of different media (audio, photo, video) with no professional mediators and little or no structuring. We argue that user-generated content creation and sharing exhibits structural properties typical for explorative rather than goal-directed search needs. This requires new kinds of search tools integrating collaborative techniques such as community-based content aggregation and filtering with visual exploration interfaces.
Author Keywords
Exploratory search, visual information interfaces, user-generated content, participatory media.
VISUAL EXPLORATORY SEARCH FOR SENSEMAKING IN UNFAMILIAR INFORMATION SPACES
Exploratory search interfaces supported by visual techniques are frequently found to perform worse than traditional keyword search with list-based search result presentation and navigation [3],[4]. In contrast, the results of our work suggest that such seemingly general weakness of exploration and visual search may be due to two main reasons: 1) task-bias i.e. generic comparison vs. domain-specific evaluations (explorative strategies may be more suitable for specific classes of application or task domains than others) and 2) too narrow evaluation metrics largely focusing only on document precision and recall, not considering the broader context of the search process and its aims (such as the topical overview of the collection or the learning effect in understanding previously unfamiliar domains).
We have investigated the development and use of visual and exploratory search techniques for specific classes of applications which require intrinsically explorative access by their nature. A concrete example is the development of methods, evaluation frameworks and tools to support a combination of explorative and goal-directed search and sensemaking in unfamiliar community spaces as part of the process of cross-community knowledge exchange [7, 8].
Seeking information in unfamiliar communities is typically motivated by ill-structured problems which cannot be solved within the user’s familiar knowledge context (e.g. in innovation processes, strategy making, interdisciplinary research or new product development). In such contexts, the information need is very ambiguous and difficult to resolve through goal-directed search. They contain tasks of inherently explorative nature and require learning about potentially relevant but unfamiliar topics and knowledge domains in order to structure the problem and identify gaps and knowledge needed for a solution, in the first place.
As a result, rather than as an information retrieval problem, in such situations the information seeking process is more correctly viewed as a sensemaking and knowledge construction process [9, 10]: an activity in which users look both for relevant information as well as for appropriate
contextualization structures in order to interpret the meaning (make sense) of unfamiliar information and its relationship to the task at hand. We have addressed this problem by developing a method and an information interface for visual exploration and multi-perspective access to unfamiliar information spaces (the Knowledge Explorer) as well as an evaluation framework for capturing user-centered search performance in terms of knowledge construction effects rather than classical document retrieval measures (document precision & recall).
The Knowledge Explorer is a multi-view visual information interface enabling the elicitation and use of personal and shared concept structures for the structuring, navigation and exploration of community information spaces from different points of view. Originally, it has been developed to support cross-community knowledge exchange mediated through information access in unfamiliar community spaces. As this application domain is characterized by highly ill-defined information needs the information seeking process is highly explorative in nature.
To address the needs of such situations the Knowledge Explorer supports a combination of exploratory visual search with goal-directed search queries and concept-based navigation (Fig. 1). It employs a combination of 2D visual document maps providing overviews of inherent semantic structure of document collections and 1D concept maps (folder trees) reflecting conceptual structures of individuals and communities of users. The document maps are based on a combination of text or multimedia-feature analysis and a self-organized neural network which groups items into clusters of semantically related content while preserving global inter-item similarity relationships [5]. The personal and shared structures are elicited implicitly from users’
bookmarking patterns and can be applied to dynamically restructure or filter a document collection, based on a
specific personal or community point of view. Document Maps and Concept Maps of an unfamiliar community allow non-members to gain a quick overview of the community knowledge structure: the main topics, documents, concepts and relationships between them. Instead of displaying community maps, using personal maps of individual users presents only a specific portion of documents which both reflect personal knowledge of the map author and are relevant for a specific information need. Similarly, applying a personal map to the information space of an unfamiliar community will classify unfamiliar documents into thematic clusters defined by the map author.
The developed evaluation framework extends document precision and recall with measures for the quality of topical structuring of user search results and the users’ learning effect of the previously unfamiliar collections as a result of the search process. An empirical laboratory experiment suggests that this framework is well-suited to measure effects of information consumption in the search process and can compensate for the bias of pure document retrieval measures [7]. The results of the experiment applying this framework to compare the use of a visual information interface (Knowledge Explorer) and a traditional search system (Google) suggest better performance of visual exploratory search in contexts characterized by ill-defined information needs and terminology problems and requiring a topical understanding of an unfamiliar information space [7, 8].
A significant difference of our solution to a number of other information and knowledge visualisation systems is that the latter focus on visualising data patterns, rather than on personal and shared knowledge structures of human users, and are conceived as tools for specialized analysis tasks (e.g. data mining, analysis of specialized document collections). In contrast, the Knowledge Explorer enables
search results
documents in zoom focus document in user focus
related concepts related concepts
Fig. 1. The Knowledge Explorer interface
the elicitation and visualisation of implicit knowledge structures of human users and focuses on its application to information exploration in unfamiliar domains. Thus, it is intended as a powerful and yet simple to use visual information access tool for normal users rather than specialized analysts. This poses special requirements on its design, requiring real-time performance, ease of use and interactive visualisation adapted to the needs of typical information access tasks in unfamiliar domains (cf.
information sensemaking). To achieve this, knowledge elicitation and knowledge discovery methods have been tightly integrated with multi-view visualization and navigation techniques in an intuitive user-centred interface.
The development of an appropriate interface and interaction design has been informed both by insights from well-known information access task models (e.g. [2], [12]) and search process models (e.g. [1], [11]) as well as by studies on knowledge construction during information seeking in unfamiliar domains (e.g. [9], [10]). In this way the vast amount of experience on developing visual information interfaces could be productively applied, while considering the particular needs of the specific exploratory search requirements of the cross-community application domain.
THE CHALLENGE OF USER-GENERATED MEDIA CONTENT
The focus on community information spaces links this work to the current phenomenon of large scale user participation in creation and sharing of content. Huge amounts of information are increasingly produced and shared between users, in a number of different media (audio, photo, video) with no professional mediators and little or no structuring.
The currently most dominant form of structuring in this context is collaborative tagging (folksonomies) which exhibits a great heterogeneity and idiosyncracy of semantic descriptors and requires large-scale user communities in order to produce usable terminological aggregations. At the same time, personal media use and sharing encompasses a number of different motivations and modes of use.
Especially the “gratification mode” is increasingly superseding the more traditionally investigated information and communication modes. In addition, end-user content creation and use have been increasingly linked to shared social contexts such as computer-supported social networks and online communities (e.g. mySpace, YouTube, Wikitravel). We argue that this change in the modality of information access together with continuing growth of personal and shared media collections fundamentally changes user search patterns and requirements for new search interfaces. The gratification mode of (multimedia) information consumption implies that the search scope is much more explorative than goal-directed, as the information need becomes even more ill-defined and difficult to express than it is commonly observed (e.g. how to formulate search terms for “nice picture”, “cool music”?). As a result, participatory media platforms tend to combine tag-based search with browsing and filtering
user feedback (implicit or explicit) such as rating, tag frequency or number of user views (e.g. Flickr, YouTube).
While this introduces some simple and easily usable structuring it also results in users being presented only with the tip of the iceberg. Since the largest part of the collection follows the long tail pattern (many small portions visited or rated by many small groups of users), the majority of items not appearing in the “top ten” remain practically invisible and very hard to find.
Furthermore, the gratification mode also changes the nature of relevance of retrieved information (e.g more relevant are items which are “more fun to see”) which makes the whole notions of precision and recall difficult to define. This implies that applying existing metrics to evaluate the suitability of existing search tools for such collections and task-domains is also becoming increasingly difficult if not downright inappropriate.
Hence, personal and shared media collections together with the phenomena of user-generated content seem to exhibit structural properties which characterize more explorative than goal-directed search patterns: the information need is extremely vague and difficult to describe, the relevance is difficult to determine, the collection is unfamiliar to the user, highly amorphous and must be structured to allow access from many different, context-dependent and often very personal points of view (e.g. moods).
OPPORTUNITIES FOR EXPLORATORY SEARCH INTERFACES
Such structural characteristics rather closely resemble our previously described work in the application domain of cross-community information seeking. The results of this work suggest that in such situations a combination of visual explorative search with goal-directed search queries and a collaborative elicitation of personal and community-based information structures can provide valuable support to the users and may outperform traditional search interfaces, both on objective measures and subjective user satisfaction [7, 8]. Our results also suggest that accessing and sharing personal points of view of unfamiliar collections can be more rewarding and effective than exploring and navigating the entire collaborative structure [8]. And we also discovered how “personality matters”: the possibility to personalize the visual presentation of one’s own collections appeals to users and is quoted as an important incentive for sharing i.e. “presenting” the personal collection to others [7]. In summary, examining the nature of current phenomena of user-generated content and personal media sharing, we find it striking how direct the resemblance of structural properties and the relationship to our experiences with exploratory search in cross-community information seeking seems. This leads us to believe that there is great potential in investigating how the existing body of experience in exploratory search, visual support and community-based content aggregation and filtering can be brought together in order to create new kinds of explorative
content collections. More specifically, we propose to extend existing social bookmarking practices with richer means for expressing personal views and creating visual representations of multimedia object collections, where spatial arrangement carries semantic meaning reflecting a user’s point of view (e.g. playlists vs. music maps).
As argued above, our experiences with facilitating information access through sharing personal visual document maps suggests that users are willing to invest effort in creating richer representations of their personal information collections in certain conditions - such as when the maps are aimed at being shared with others (self-presentation, reputation and hedonic factors) or when gratification is the primary motivation of information use.
As demonstrated in our previous work, user bookmarking patterns can be used to construct personalized views of information collections presented in form of visual overviews which map similar items close to each other, based on examples previously classified by the user.
Similarity measures can also be defined based on document collocation patterns (e.g. the dice coefficient [6]) by communities of users (which have declared similar interests or opted for explicit community membership). This is a well-known technique which can avoid the problem of high idiosyncracy of user tagging patterns and actually be used to infer relationships between different classes of semantic descriptors for mutually related items. As a side-effect, tags signifying different but possibly related contexts of use (e.g.
“easy listening” and “jazz” used for the same music) can be put in relation to each other.
Accordingly, user-induced “personalized media maps” can then be used as visual semantic templates for guiding exploration of otherwise unstructured collections, from different personal points of view. They can function as personalized templates for contextualizing keyword searches or conducting implicit search through “filtering by example” in situations where the information need cannot be easily or unambiguously expressed. And they can be aggregated across a group of related users (e.g. members of a community or a specific thematic social network) to provide visual means for browsing areas of the collection corresponding to shared interests of a group of users.
Finally, since they represent rich semantic contexts, such maps can be used to contextualize keyword-based search results: identifying the most relevant personal collection for a given search query provides a visual context within which the query results can be presented to the user (classified and spatially positioned into topical areas defined by the map author). In this way, related items can be discovered by visual exploration even when they greatly mismatch the original query: either by serendipitous browsing or by being spatially collocated with some of the retrieved documents but based on other attributes not targeted by the query.
Thus, by facilitating the creation and sharing of user-induced visual semantic templates we could overcome the
limitation of current solutions for explorative access to user-generated media content collections which tend to provide rather restricted “tip of the iceberg” views. Instead, combining explorative visualization with personalized classification and community filtering seems a promising way for enabling rich visual exploration of user media collections based on different granularities of access (from personal to community level) and different facets of contextualization (e.g. different maps representing different selection criteria, based on varied needs and use contexts).
REFERENCES
1. Bates, M.J. The design of browsing and berrypicking techniques for the online search interface. Online Review, 13(5):407.424, 1989
2. Belkin, N.J., Stein, A., Thiel, U.: Cases, Scripts and Information Seeking Strategies. On the Design of Interactive Information Retrieval Systems. Expert Systems with Applications, Vol. 9(3): 379--395, 1995.
3. Chen, C., Yu, Y. Empirical studies of information visualisation: a meta-analysis, Int. Journal of Human-Computer Studies, Vol. 53,No.5, 851-866, 2000
4. Visualization of search results: a comparative evaluation of text, 2D, and 3D interfaces Sebrechts, Cugini, Laskowski, Vasilakis and Miller, Proceedings of SIGIR 99, Berkeley, CA, 1999.
5. Kohonen, T., Kaski S., et al.: Self Organization of a Massive Document Collection, IEEE Transactions on Neural Networks, Vol. 11, No. 3, May 2000.
6. McGill, M.et al.:An evaluation of factors affecting document ranking by IR Systems,” Project Report.
Syracuse University, School of Information Studies, Syracuse (NY), USA, 1979
7. Novak, J. Collaborative Knowledge Visualisation for Cross-Community Knowledge Exchange, PhD Thesis, University of Duisburg-Essen, Germany, 2006
8. Novak, J. Helping Knowledge Cross Boundaries:
Using Knowledge Visualization to Support Cross-Community Sensemaking, Proc. HICSS-40, Hawaii, 2007
9. Qu, Y., Furnas, G.: Sources of Structure in Sensemaking. Proc. of ACM CHI 2005, Extended Abstracts and Applications, Portland, OR, USA, 2005 10. Russell, D. M., Stefik, M. J., Pirolli, P., Card, S. K.:
The Cost Structure of Sensemaking. Proc. of ACM INTERCHI'93, pp. 269-276, 1993
11. Saracevic, T. The stratified model of information retrieval interaction: Extension and applications. In Proceedings of the American Society for Information Science 34 (1997), 313–327
12. Shneiderman, B. The Eyes Have It: A Task by Data Type Taxonomy for Information Visualisation.
Technical Report 96-66, Institute for Systems Research, University of Maryland, 1996