In this paper we present a general purpose solution to Web content perusal by means of mobile devices, named Social Context-Aware Browser. This is a novel approach for the information access based on the users' context, whose aim is to retrieve what the user needs, even if she did not issue any query. Our solution is built upon a social model that exploits the collaborative e orts of the whole community of users to control and manage contextual knowledge, related both to situations and resources. This paper presents a general survey of our solution, describing the idea and presenting an implementation approach.
Context-aware computing is a computational paradigm
that has faced a rapid growth in the last few years,
especially in the eld of mobile devices. A key-role in this new
approach is played by the notion of context, that is roughly
described as the situation the user is in. This concept
encloses important information that could be used to a ect the
capabilities of mobile devices, adapting them to the user's
needs. In particular, contextual data can be used to
predict the user needs and to seek and retrieve information,
thereby reducing the complexity of the user-device
interaction and providing the right information in the right place
at the right time. From this point of view, because of the
huge amount of contextual information and its
heterogeneity and uncertainty, the mobile and context-aware
computing environments represent a new challenge for Information
Retrieval (IR). The combination of IR and context-aware
computing has been named context-aware retrieval [
These considerations guided us towards a new approach
to Web contents production and fruition, where contextual
data are exploited to capture the dynamic nature of the user
needs, of the information available, and of the relevance of
this information, typical of a mobile user in the real world.
This approach is named Social Context-Aware Browser and
its novelty is threefold. First of all this is a new radical
approach that aims at discovering \the query behind the
context": to retrieve what the user needs, even if she did not
issue any query [
This paper is structured as follows. We rst brie y survey related work (Section 2), presenting the Context-Aware Retrieval eld and introducing the main ideas behind Web 2.0. We then describe our solution (Section 3), presenting a general survey, the main ideas, and an implementation approach. In Section 4 we present a brief discussion and nally we draw some conclusions and we present future work (Section 5). 2. 2.1
Context-Aware Retrieval (CAR) is an extension of
classical Information Retrieval (IR) that incorporates the
contextual information into the retrieval process, with the aim
of delivering information to the users that is relevant within
their current context [
Typical CAR applications present the following
characteristics [
An example of CAR system is the Ubiquitous Web [
With Web 2.0 [
Social bookmarking is a method for organizing, searching, and managing documents of interest among users. In a social bookmarking system, users save links to documents of interest in order to remember or share them with the community. Social bookmarking is strictly related with the concept of folksonomy, that is the practice of annotating and categorizing content in a collaborative way, by means of informal tags. Folksonomies, that is a portmanteau of folk and taxonomy, allow users to easyly and informally descrive documents and content. This represents a powerful combination that has gained popularity as it allows a more natural and simpler management of the knowledge. The use of freely choosen categorizations and the collaborative aspect in fact allow also non-expert users to classify and nd information. Folksonomies and social bookmarking for example are used in well-known Web 2.0 systems like Flickr1, Youtube2, Del.icio.us3, etc.
Folksonomies however are criticized because the lack of
terminological control could lead to unreliable and
inconsistent results [
The Social Context Aware Browser (sCAB for short) [
The sCAB acquires information related to the user and the surrounding environment, by means of sensors installed on the device or through external servers. This information, combined with the user's personal history and the community behaviour, is exploited to infer the user's current context (and its likelihood). In the subsequent retrieval process, a query is automatically built and sent to an external search engine, in order to nd the most suitable Web pages for the sensed context and present them to the user.
As current models for context-awareness are too limited for very general applications like the sCAB, this approach brings new social models for CAR that exploit the collabo1www.flickr.com 2www.youtube.com 3www.del.icio.us.com rative e orts of the community of users. The community, in fact, is encouraged to de ne the contexts of interest, share, use and discuss them, associate context to content (web pages, applications, etc.), to have a dynamic and more usertailored context representation and to enhance the process of retrieval based on users' actual situation.
In particular users can freely interact with resources and can de ne that a resource is useful (or not adapt) to their current context, can associate resources to particular contexts, can explicitly de ne the context their are in, and nally can browse resources relevant for their current context. 3.2
3.2.1
We represent the context as a folksonomy. Each tag is
banally a keyword or string of text and represents a single
contextual value [
Concrete tags: represent the information obtained by a set of sensors. These information can be read from the surrounding environment through physical sensors (e.g., temperature sensor), or can be obtained by other software (e.g., calendar) through logical sensors. Concrete tags that directly refers to sensors values are represented using the triple tags notation that are tags that uses a particular syntax (namespace:predicate=value) to de ne extra information.
For example, geo:longitude=12.456 is tag for the geographical longitude coordinate whose value is 12.456. Other concrete tags, can be automatically obatined by the sensed values (e.g. afternoon, summer, ...).
Abstract tags: represent the high level contextual information that are freely associated by the users to the concrete contexts, in order to detail their context description. Some examples are: home, shopping, etc. The di erence between the two categories is faded since the contexts cannot be unambiguously assigned to one or the other category. However this partition is helpful in order to distinguish the low level information coming from sensors and the high level contextual information intoduced by users.
The user context is a \cloud" composed by an unde ned number of concrete and abstract tags (Figure 1).
In the sCAB conceptual model [
The other four operations are strictly related to the user interaction: the main two are de nition and annotation (Figure 3). The de nition is used to manage the contextual information and it is performed when a user directly de ne her context, or when she provides contextual tags during the annotation of a resource. In particular, this operations manages the associations between concrete and abstract tags, and the strength of their relationships. The annotation on the contrary is used to manage the association between contextual tags and resources and it is performed when the users link resources to particular contexts. We can imagine a user at a park with her dog: she wants to associate to her context a particular Web page teaching dog training. For this reason she bookmarks that resource with the contextual tags \out dog park sunny train". Doing so, rst the added abstract tags are related to the sensed concrete tags and for all the users with a similar concrete tag cloud, these abstract tags (or part of them) can become part of the their context representation. Second, that particular Web page is enhanced with all the tags, and it will be automatically proposed to users every time they will be in a similar context.
As the users are the main actors in the process of context
de nition and resource annotation, problems related to the
quality of context and resources are likely to appear. To
cope with this problem we propose the adoption of a social
evaluation/reputation mechanism. We exploit the ideas
presented in [
Concrete and abstract tags, and resources are the main elements in our implementation model. Concrete tags, as output of sensors, are exploited to retrieve the most relevant abstract tags, and in the same way all the tags are exploited to retrieve the most relevant resources.
In the following sections we show an implementation proposal and how the di erent operations in the model have e ect on the system, from a low level point of view. 3.3.1
We exploit two indexes. In the rst one, called contexts index, abstract tags are indexed over concrete tags, while in the second one, called resources index, resources are indexed over the set of all tags (both concrete and abstract). The proposed approach is community based, thus the indexes and the inferential system are managed by remote servers and not stored on the mobile device. Since the approach is similar for both the indexes, we are going to show just the rst one.
The contexts index is a matrix that describes the frequency of abstract tags over the concrete ones. Each column corresponds to a concrete tag, and each row corresponds to an abstract tag. Each entry in the matrix has three values (Figure 4):
Uij : represents the user that has associated the abstract tag i to the concrete tag j rst; Sij : a score that de nes how relevant the abstract tag i is for the concrete tag j. This value is in the interval [0; 1];
ij : steadiness value that de nes how steady is the association between the abstract tag i and the concrete tag j.
a1 a2 . . .
Intuitively, since not all the abstract tags can be related
to all concrete tags, the proposed index will be a very sparse
matrix. At the same time, because of the very high number
of both concrete and abstract tags, the index can assume
very huge dimensions. However a lot of research is being
performed on indexes designing and analysis, also in the
CAR eld [
In our approach two values are associated to each user and they de ne the goodness of the user in working with contextual information:
SUc : a score that de nes how good the user is in associating concrete tags to abstract tags; SUr : a score that de nes how good the user is in associating resources to contexts; As previously, we are concentrating only on the management of values related to concrete and abstract tags, since the approach is exactly the same working at the higher level of tags and resources.
Every time a new relation between abstract and concrete tags is created with a de nition (\ lling a hole" in the index), the user who performed the operation is associated to that relation. Then on the basis of how the community interacts with those contextual information, the user's score will be update. It is calculated as follows: for each association among tags ij performed by the user U , SUc corresponds to the mean of the products ij Sij, where max is the max max steadiness value in the index.
New associations have a low steadiness value, thus their score, as their have not steadied yet, will have low in uence on the user's score. Good associations will have high score and steadiness values, and they will re ect on high users' score. In the same way, low users' scores are due to bad associations between contextual tags. Since Sij 2 [0; 1], also SUc 2 [0; 1].
In this approach, for simplicity, only new associations between tags are considered for the computation of the users' score. An extension could consider all the existing associations. In this way a user is \good" because she de nes good new associations and because she exploits existing good association. 3.3.3
The proposed indexes are not static, but the values related to the association between concrete and abstract tags and resources are continuosly updated, based on the interaction of users with resources in context.
With every de nition operation the values in the contexts index are updated according to the following system (for the values in the resources index with the annotation operation the approach is similar) : ij(ti+1) = ij(ti) + SUc (ti) ij(ti)
Sij(ti)
SUc (ti) v = Sij(ti+1) =
ij(ti+1) v if v > 0 0 otherwise where ti represents a discrete time instant and ti+1 the subsequent time instant.
While the score is a value in the interval [0; 1], the steadiness is an always increasing value. The higher the steadiness of an association is, the more stable the association is, and then the lesser e ect each update operation will have. The user's score is exploited for the update of the values in the index. It can both increase an association, or decrease it (e.g. a user removes a tags from his context). The higher the user's score is, the more e ective the update operation will be. This means that good users have more in uence on the system than bad users. Finally, is a parameter greater than 0 and it is used to weight the user score: operation performed explicitly by users (inclusion or removal of abstract tags) have more e ect than implicit update performed automatically based on the interaction of the community with the resources.
The inference and retrieval operations works respectively on the rst and second index, but they are similar, thus in the following we are explaining just the inference one.
The approach is the following: 1. starting from the concrete tags in input, we consider only the set of abstract tags that have been associated at least with one of the concrete tags; 2. for each abstract tag we compute a rank value, to dene an order of relevance for the abstract tags; 3. in order to limit the number of retrievd tags, we retrieve the abstract tags whose rank value is higher than the mean of all rank values.
The rank value is computed following an adapted version of the tf.idf weighting scheme. In particular for each considered abstract tag ai we have:
A = Pcj ij
Sij, for each sensed concrete tag cj B = jCj , where jCj is the total number of jfc : ai 2 cgj sensed concrete tags, and jfc : ai 2 cgj is the number of concrete tags to which the abstract tag ai has been associated; rank value = A B , where ; are parameters exploited to weight the di erent values.
Some considerations can be drawn. First, more are the concrete tags in the current context to which an abstract tag is associated, the higher will be its rank value. Second, abstract tags with high score and steadiness will have an higher rank value. Third, abstract tags related to particular sets of concrete tags will have an higher rank value than very general ones that are associated to an high number of concrete tags (high frequency).
In addition, starting from this basic approach, we can enhance the rank value computation exploiting other information. For example a reasonable idea is to weight the tags based on their age in the user's context representation, giving more importance to the newest tag. In this we enhance the importance of new contexts.
Although the conceptual ideas are clear, the implementation approach we propose is in an initial stage of de nition. We suggested a possible solution, but several are the ways to re ne it and several are the algorithms to be exploited. For this reason the evaluation hold an important role in our work: since di erent alternative solution exist, it is important to evaluate them and compare their e ectiveness.
Even if the knowledge related to the whole community is exploited to infer and re ne the current context of single users, the proposed model di erentiates the personal from the community level, giving more importance to the rst one. For example if a user annotates a situation as \play", she is considered to be in \play" context, even if most people annotate the same situation as \work". On the contrary, if a user is for the rst time in a situation (e.g. location never visited), her context is re ned just with the information from the community. Considering the previous example, as most people annotate the situation with \work", the user is considered to be in \work" context.
In the last case, the assumption performed by the system in order to provide the user with relevant resources could be wrong. However this is not a problem. Since we are working with people, it will be hardly possible to provide results that totally satisfy each user, due the intrinsic di erence of views and needs in a community. Rather our solution aims at and averagely good behavior.
Talking about the indexes, we have seen how the related information are changed dynamically based on community interaction. However this is not the only possible approach. We can imagine complementary approaches that can support the community statistical one. For example, we could use some geographic gazetteer for associating geonames to geographic coordinates provided from the concrete tags, so as to reinforce the rank of associated abstract tags that contain the same geographic names or names of close localities. The geonames could be useful also for retrieving more relevant resources, those containing the geonames ore close geonames.
In this paper we have presented the Social Context-Aware Browser, a general purpose solution to Web content perusal by means of mobile devices. The sCAB is a novel approach for the information access based on context, where the community of users is called to manage the contextual knowledge, both related to situations and resources, through collaboration and participation. In particular we presented a general survey, the main ideas, and an implementation approach.
As future work we aim at implementing a prototype of the
proposed system, and, in particular, we suggest a multistage
approach, where implementation and evaluation processes
will proceed hand in hand. As rst step we want to exploit
benchmarks to evaluate detailed implementation solutions,
like, for example, di erent algorithms to assess the relevance
of tags for situations and resources. After that, we plan to
apply an IIR evaluation methodology, involving users in a
controlled environments, following the ideas presented [
The authors acknowledge the nancial support of the Italian Ministry of Education, University and Research (MIUR) within the FIRB project number RBIN04M8S8, and the region Friuli Venezia Giulia. This research has been partially supported by MoBe Ltd. (www.mobe.it), an academic spino company specializing in software for mobile devices.