In this work we report about the automatic creation of dynamic user pro les that combine personal data from diverse social networks with context information recorded by IYOUIT, a mobile community service in the eld of context awareness. Data mining algorithms are applied to learn about users and their preferences over time. Rich, semantically annotated Friend of a Friend (FOAF) pro les are created on demand, and thus always re ect the latest level of information. Existing extensions and potential enhancements of FOAF are discussed.
In order to provide truly personalized services, having knowledge about the meaning and the correlation of personal information streams is inevitable. In the following, we present our attempts to interpret diverse types of information, either coming from sensing the user's vicinity via the mobile phone or retrieved from connected Web 2.0 services. As a result, rich user pro les can be deduced to explore social interaction in the virtual world. This paper summarizes rst results of newly established research aspects in IYOUIT that focus on context-based user pro ling and the adoption of common semantic representation formats. 2
With IYOUIT's mobile application personal information including digital location traces, status updates or communication patterns can be automatically recorded, processed and in turn made available to the user on the mobile handset or on the Web. From localized service o erings to real-time presence updates of registered buddies, IYOUIT o ers a broad range of bene ts to the mobile user. Most personalized service o erings require user pro les that hold information about basic demographics and application speci c preferences. To facilitate and substantially improve the automatic creation of such pro les, data mining algorithms are applied. Here, the general aim is to learn about the user, his habits and liking to increase the personal pro t in participating and contributing to the community.
IYOUIT builds upon a distributed component infrastructure to allow for exible service deployments. Each distinct type of context information is therefore processed in a corresponding network component, in the following referred to as ContextProvider, and is made available to clients and other components through a rich API. Whenever possible, a ContextProvider is meant to abstract from raw data sources to qualitative information. In some cases, however, semantic references cannot be easily made, given the fact that some pieces of information, e.g. manual presence updates, may not be categorized subsequently.
Based on continuously recorded location traces, the LocationProvider
applies clustering techniques [
Since each ContextProvider may also encapsulate external data sources as well as 3rd party services, data provided by popular Web 2.0 services is in turn utilized within IYOUIT. To give an example, Last.FM, a community service that lets people socialize around music, o ers an API to access a registered user's pro le as well as music favors. This pro le information can be enriched with automatically recorded usage patterns on the mobile handset. This way, not only the most popular artists or genres of a user are known, but also the context in which a certain type of music is preferred. Having information about the time of the day when a user is usually playing music in addition to associated activities such as commuting or being at home allows for signi cantly enriched music pro les. 2.1
Even though distinct, context-type speci c user pro les are available in distributed components, the management and enrichment of pro les across multiple context types is accomplished by one central component - the Pro leProvider. In regular time intervals, all published pro les are harvested from registered components to provide a single endpoint for the retrieval of pro les for services building on top of the framework. The Pro leProvider's main task, however, is concerned with the cross-context analysis of personal data, as summarized in Figure 1. So-called context summaries, an aggregated variant of the actual data record, are retrieved periodically from all available components. The subsequent procedure identi es a uniform weighting of the context summaries retrieved, to compare di erent types of information with regard to their importance in the given time interval. References to the original information residing in the respective network component as well as its measured signi cance values are stored for further assessment. This context retrieval process is executed in periodical time intervals, eventually resulting in multiple snapshots of a concept's signi cance distribution over time. A concept stands for any atomic piece of information being considered, for instance information about staying in a personal place or updating ones presence information. Semantic associations that may have been computed within the respective ContextProvider, e.g. references to the place ontology, are maintained for subsequent reasoning tasks. The signi cance distribution of a concept results in trends of increasing, decreasing or steady importance.
Even though a certain trend of a given concept may indicate its relevance over time, the overall importance of a concept in a user's digital trace cannot solely be characterized by its presence in only one context stream. In addition, the degree of correlation with other concepts of di erent types and their respective co-occurrence considerably adds valuable meta-information about a concept's signi cance in general. As soon as a concept is being considered relevant, its ..
..
DB + r e d i v o r P e l fi o r P
temporal correlation with other context categories is computed. The computation of such context correlations is accomplished within the ContextCorrelator. Here, the temporal co-occurrence is the deciding factor whether or not two or more concepts are marked as correlated. To improve this co-occurrence measurement, each distinct type of context has been annotated with a validity attribute. The validity attribute has been assigned by the respective ContextProvider and is meant to quantify a concept's temporal duration. To give an example, an actual weather reading may have a validity of one hour, whereas a single location measurement may only be considered valid for 10 minutes.
The ContextCorrelator does not only identify correlations in multiple
context streams but also assigns a weighting factor to each correlation pair. Based
on these correlations and their corresponding weighting factors, association rule
learning algorithms are applied subsequently. The Pro leProvider incorporates
the WEKA toolkit [
activity="activity.owl#listening_to_music" and genre="music.owl#jazz" 2.2
The prime advantage of having rich user pro les in place are better, personalized service o erings. Usually, a community service commits itself to connecting users across explicitly given social relationships to leverage interaction and information exchange. To do so, the computed pro les and deduced patterns are further utilized in IYOUIT by applying a matchmaking mechanism, meant to identify and weight similarities in user pro les within the community. Users with similar hobbies, interests or preferences can thus pro t from being implicitly connected. Input parameters for the actual matchmaking algorithm include a concept's current signi cance, its temporal distribution (trend) as well as its correlation degree. The latter can be seen as another indicator for a concept's overall importance among a given set of concepts. The weighting factors for each input parameter may vary depending on the actual application. In case the latest state of a concept's importance is the key factor, its signi cance value can be seen as the deciding factor. On the other hand, long-term considerations can be stressed by weighting a concept's temporal distribution accordingly.
So, in order to provide such a matchmaking algorithm for nding other users with similar interests, for instance as part of an online community portal, several other aspects with respect to the actual visualization of pro le information in general and a corresponding query interface in particular need to be considered. Since the resulting user pro les are meant to be edited and reviewed by the user, the actual requirements on the data visualization are manifold. For one, the user should not be overwhelmed by the amount of information being displayed while, at the same time, the dynamic aspects of the pro le data should be considered, too. On the other hand, the user should also be able to browse through pro le information of other users in case a certain match has been found.
The most common way to visualize (mainly unstructured) data on the Web is by utilizing tag clouds. With tag clouds, the importance of concepts can be visualized by a tag's overall font size. However, only a very limited amount of meta-data is encoded within the cloud, preventing important information from being visualized. On the other hand, the advantages of a tag cloud based visualization are its wide adoption, ease of use and applicability to diverse domains. To overcome the limiting factors while still maintaining some bene cial aspects, we came up with a more advanced visualization paradigm, still borrowing the basic concept of representing data in a way similar to tag clouds. In short, the signi cance distribution of a tag over time is visualized by changing its font characteristics (e.g. by using di erent font sizes) within the representation of one tag. So in case the signi cance of a tag has decreased signi cantly, the font size of the distinct tag decreases, too, and vice versa. As a result, the tendency of the observed signi cance of multiple tags can be visualized in one (static) cloud, and thus enables a better identi cation and selection of tags not only based on their current signi cance, but their signi cance distribution over time. In addition, being able to visually identify trends of single data items or groups of items, it becomes possible to not only better understand the provided information but also to browse to historic samples of the data as recorded in the past. Once a certain tag has been selected, distinct parts of this tag that represent historic samples of the data set can be picked. As a result, the tag cloud is recomputed given the chosen temporal constraints. Likewise, through visualizing a tag's signi cance distribution over time, a trend-based browsing through the given dataset becomes possible.
The actual font sizes used for a given tag and its signi cance distribution can be computed by assuming that the number of distinct sections (or letters) of a given tag also restricts the number of signi cance measurements considered. This trend visualization can be described by
k 8i=0fi = Fmin + s i (Fmax Smax
Fmin) with fi being a distinct fonts size, Fmin and Fmax representing minimum and maximum font sizes, si being a distinct signi cance measurement and Smax representing the maximum signi cance measurement. Apart from those temporal aspects, existing tag cloud representations do often not allow for visualizing the degree of correlation and therefore lack another suitable di erentiating factor between tags in one cloud. As a result, navigating through the dataset and restricting a given query along the lines of correlated tags is not possible. Our proposed visualization and interaction techniques tackle this constraint in visualizing the correlation factor between tags with either a distinct grouping, colorcoding or distance indication. Figure 2 depicts a screenshot of our Web-based pro le browser that implements the aforementioned visualization characteristics. By visualizing the correlation of tags by, for instance, more than one color, it becomes possible to di erentiate between concepts representing a premise item set from the ones representing consequence attribute/value pairs.
Through the selection of one distinct tag, its signi cance and correlation
with other concepts can be assessed. At the same time, the selected tag's
metainformation is used to nd matching descriptions in other user's pro les. The
resulting weighted list of users is retrieved and displayed in a separate drawer
on the left hand side of the pro le browser. Here, the similarity degree of the
pro les determines the ranking within the result set. Once another user's pro le
has been selected, the interactive tag cloud displays the selected user's pro le. By
browsing through di erent (anonymous) pro les, patterns found within the cloud
can be explored to draw meaningful conclusions about the respective user. The
pro le browser as depicted in Figure 2 represents a prototypical implementation
of the proposed visualization paradigm and will thus be further re ned with
respect to its adoption within the deployed service framework. The IYOUIT user
community is invited to try out several aspects of the automatic user pro ling
and the pro le matchmaking to provide valuable feedback on the practical use
and general perception.
8"()$.,52)4"0/2)*+,-./)
+/*+/0/'#"6,')
!"#$%&'()*+,-./0)
Even though some semantic references are present, the underlying format of the
pro le data mainly builds upon a hierarchical XML representation shared among
all IYOUIT components, including meta-information such as a concept's
correlation or signi cance distribution. This allows for e ciently processing large
amounts of data with common technologies while, at the same time,
higherlevel reasoning can still be accomplished by references to underlying
ontologyconcepts. However, a complete semantic foundation of the pro le data would be
bene cial in some respects. For one, a standard vocabulary for describing user
pro les makes data exchange across service providers far easier. At the same
time, a growing number of tools in the eld of the Semantic Web provide
support for enhanced query mechanisms across distributed knowledge sources [
The core of each FOAF pro le is comprised of the social network of the
respective person. With the foaf:knows property, other people and their FOAF
pro les can be interlinked. As stated in the vocabulary speci cation, foaf:knows
is rather vague by design. Some applications, however, require a more
negrained representation of social relationships between people. IYOUIT is an
application of the latter, and provides means to express a social network by
well-de ned properties in a social ontology [
Another rather small vocabulary that is being used to express information about a user's visited countries is VISIT.7 Through IYOUIT's worldwide adoption of people who tend to travel frequently, this addition is most welcome to represent at least one aspect of their travel activity. By constantly observing a user's location traces, trips can be automatically detected by computing the distance between subsequent measurements and the resulting speed approximation. Being able to automatically recognize trips, a user's pro le can be updated instantly and without any user interaction.
In fact, most context data available within the IYOUIT framework is primarily concerned with real-time presence information about a user's current activity. Even though the principle use of FOAF is focused on rather static pro le and contact information, the idea to also represent highly dynamic and frequently changing data seems to be reasonable as well. However, to actually state the timeliness of information, a timestamp must be present along with the data attribute and its value. One existing approach that utilizes dc:date attributes is the so-called MeNow vocabulary.8 Basically, MeNow aims at describing the 6 http://vocab.org/relationship/ 7 http://purl.org/net/vocab/2004/07/visit 8 http://crschmidt.net/foaf/menow/ current status of a person. Its properties allow for expressing that a user is currently reading a book, the type of music he is listening or the websites he is browsing, amongst others. Also worth noting is the possibility to express the physical proximity of other users and their online activities. Therefore, MeNow addresses the growing demand for vocabularies that express the timeliness of information, as required for applications dealing with frequently changing information. An excerpt of an automatically created IYOUIT user pro le is shown in the Appendix, including applicable real-time presence information encoded with the MeNow vocabulary. 4
The FOAF vocabulary is the semantic representation format of choice for
expressing social data in online networks. It has been explored in a broad variety
of research areas with recent interest stimulated by the Linked Data approach
[
Likewise to the semantic representation of data, the visualization of
information on the web is a well-established research topic, too. In [
The steady increase in the availability of semantically annotated knowledge sources across the Web and the development of tools making use of that information drives the ongoing trend towards a widespread, practical use of semantic data. Even though FOAF has been around for almost a decade, it has seen recent uptake through Google's attempt to crawl publicly available FOAF pro les and Yahoo's SearchMonkey9 and its support for structured data, including RDF.
Typically, the information represented with FOAF can be seen as rather static, involving less frequent updates. However, the way people tend to use tools and services on the Web and in turn their personal re ection in social communities has changed considerably over the years. Frequently updating ones presence in Twitter and the provision of personal information is common practice for a lot of people. With IYOUIT we aim at simplifying this process through the combination of context information with personal pro le data. Rich user pro les are automatically created and published, always re ecting the latest level of information. The timeliness of the information and the rich descriptions of a user's preferences are meant to provide the basis for novel, user-centric services.
In utilizing FOAF as the external representation format in combination with
the use of cool URIs [
A future aspect of our work is to evaluate the use of exible semantic
representations, like (extended) FOAF, also as basic data structures internally within
IYOUIT components. Preliminary results on realizing an IYOUIT social network
visualization on a FOAF/RDF substrate via corresponding SPARQL queries
demonstrated greater exibility and improved robustness. The common
representation format allows to interlink distributed social graphs through the
interpretation of foaf:holdsAccount, owl:sameAs and rdfs:seeAlso triples in
a generic way and thus allows for browsing friendships across social networks
(almost) for free. Web 2.0 accounts that do not provide FOAF descriptions like
Flickr, Twitter, Last.FM or DBLP [
<dc:title>X-Stream</dc:title> </foaf:Project> </menow:workingOn> </menow:Status> </menow:hasStatus> </foaf:Person>