The possession of mobile devices eased a lot our communication habits: contacting somebody in almost any situation is becoming more and more easier. However, this implies the fact that unexpected interruptions can occur that may cause discomfort to the user. A typical example is the case when an important meeting is interrupted by an unexpected phone call, which is not important to consider in the given context. The need for an e cient control of unexpected incoming communications is immediate. This paper proposes the e cient handling of interruptions by providing a new reachability preference model that considers two dimensions for a reachability preference: (i) the currenet situation of the user and (ii) the social category of the communication partner. We propose a mechanism to build user preferences that specify the level of interruptibility during an activity, situation or event. The proposed system leverages semantic technologies for modeling the user and data mining techniques for populating the user model.
Today, as smart phones become a daily companion for most people, user
modeling must address several new challenges. In our point of view, one of them is
the issue of how to provide meaningful assistance to end-users in
communication needs. One domain is the reachability of the user in a given situation. We
de ne reachability as the preference of the end-user to be interruptible (by a
given service, like a phone call, sms, IM, noti cation) at a given time or time
interval. This choice for the de nition is mainly based on [
From the user's point of view, assistance in reachability is an e cient ltering
of interruptions to his current state of mind or activity. As pointed out in [
For interoperability purposes, our user model is based on semantic
technologies (OWL as description schema, SPARQL query language for information
retrieval). This choice is based on current research initiative to leverage semantic
technologies for a richer user pro le representation. Several user pro le models,
like GUMO (General User Model Ontology) [
The rest of this paper is organized as follows. Section 2 presents the ontology for the user pro le model. Section 3 presents an upper-view of the framework. Future enhancements of the system and conclusions are presented in Section 3.
The objective of this user model is to allow end-users to have personalizable preferences in terms of reachability that are automatically activated. Preferences are personalized according to three dimensions: the current situation of the user , the service (phone, IM, SMS, Mail) and the social network category (friend, family, professional etc.) of the person who is behind the incoming interaction. We divide the user model into submodels (Figure 1) (Concept: SituationalProleSubset ). A submodel contains user preferences that are activated when that speci c situation occurs (Concept: Situation). A situation is described by several contextual information (Concept: Context ), composed of the Physical Context of the user (Location, Environment etc.) and the context relative to the user itself (Activity, Agenda, Tasks etc.). Each submodel contains the reachability preferences for the given situation. A preference concerns a Service (Phone, SMS, IM etc.) and a Social Network Category (Friend, Family, Professional etc.) and contains the noti cation modality for the Service in case it is activated by a contact belonging to that category. An important aspect of this model is the categorization of the address book into categories (SocialNetworkCategory). This needs a manual intervention of the user but the bene ts are considerable.
To illustrate the model, we consider an instance with two sub-pro les : (Home Pro le and O ce Pro le). Each sub-pro le is activated when the given situation occurs (e.g. the user is at home or at work). The Home Pro le is associated to the At Home situation, which has the following context: the location of the user is Massy (a French town near Paris) - inferred from the Cell Tower Identi er from the mobile phone- , it is evening (Time), the user is with his son (Bluetooth peer detection) and the user is preparing to watch a movie(Agenda). Here each value is an instance of a context concept (Location, Time, Environment, Agenda). In this pro le, the user has two preferences, the Home Preference, which de nes that when friends (Carla) and family members call, the phone will ring. The other preference is the O ce Preference, which consists in the fact, that if an o ce colleague tries to call, he/she will receive a message: "I am currently available on E-Mail.". The second sub-pro le is the O ce Pro le with 2 preferences. A Work Preference states that when the manager calls, the phone will ring. The Family Preference states the same thing about important family members. However, the Friend Preference states that if a friend calls (Carla), the phone will only vibrate. 3
The reachability framework contains tools for both lling the user pro le with information from di erent data sources and for retrieving this information. These tools are separated in a front-end (information retrieval) and a back-end (prole lling, data sources). Both components will be described in the following paragraphs.
The front-end contains SPARQL query templates that are instantiated to retrieve the noti cation modality from the user model. Such a query template contains three parameters in input: the situation, the service and the category of the incoming caller and the result is the noti cation modality for the service. SPARQL was the optimal choice for this, since the user model has an OWL representation.
The back-end of the framework contains the necessary mechanism for the automatic pro le construction and evolution. Our choice was for a exible user model in terms of preferences. An argument for this is the fact that neither situations (the location for the Home situation can change), nor social network categories are stable (a friend can become a family member, professional relationships can change etc.). The framework uses a technique that combines data mining and machine learning in order to learn and update context of the di erent situations of the user. The context of a situation is the primary source to trigger a preference, since is contains machine-interpretable raw-data (a location coordinate, a Bluetooth address). It is important to stress out that this algorithm is only able to nd general or regular situations of the user (habits), like Home, Work, Shopping. Another data source is used to handle social events, that occur occasionally, like going to the cinema on Friday night. A situation has a rich contextual description. In the current state of the framework, we are able to capture phone cellule identi ers for location and the list of available Bluetooth peers, that give information about the social environment of the user (single or in an open space): Context Snapshot = fTime of the snapshot, list GSM Cell ID, list Bluetooth Peersg.
In the back-end of the framework, we use a multi-layered algorithm that
extracts habits of the user to learn situations and populates the ontology
frontend with the learned situations. Situations are strongly related to user habits. A
habit is a frequent pattern or item in the user's context or activity. A frequent
pattern is an association of multiple context dimensions that frequently occur
together. The signature of a situation is a list of context dimension instances,
that indicates in which context the situation occurs for the user. In this way, if
we log the user context, nding situations in the context log is a problem similar
to nding frequent patterns in any kind of data log and can be solved with
data mining techniques [
(1) The second step in this layer is to combine elements from FrequentItems and generate association rules. An association rule of two or more frequent items creates rules with high con dence (superior to a MinConf threshold) in that if the LHS (left-hand-side) is present in a context snapshot, the RHS (righthand-side) will also be present. Con dence can be interpreted as an estimate of the conditional probability of RHS, knowing the probability of LHS, that is the probability of nding the RHS of the rule in transactions under the condition that these transactions also contain the LHS.
F requentRules = f[(A ) B);
Supp(B)
(2) The result of the rst layer is a set of associations between context dimensions in form of rules. These rules encode the signature of a situation candidate with two measures, the support and the con dence. The second layer is responsible of classifying these rules. The classi cation Layer allows merging similar situations and erasing those that did not occur for a long time. This layer allows the method to be incremental (the database can be erased after extraction, which is very important because of the limited storage space on a smart phone). First, a merge of the LHC and RHC of association rules is realized, creating a signature from the rule. A situation class candidate is created for each signature. For each signature, the algorithm computes the distance to all other (existing already or extracted) classes, except the signature itself. Distance is computed by the number of co-occurrences in the two signatures. Then, the signature is uni ed with the class that is the most similar. A local age parameter is associated to each situation class. This parameter allows to erase classes that no longer appear.
The result of this classi cation step is a limited number of situation signatures, which have the following properties: each context dimension respects the minimal support threshold, the signature is a union of very similar association rules and contains strongly correlated elements since this union is based on association rules respecting the minimal con dence level and this signature contains all possible context dimensions that have a high probability to occur when the user is in the given situation.
As mentioned before, this algorithm is able only to nd regular situations. The detection of exceptions (in this case occasional events) is a big challenge in any data mining approach. We consider the user calendar (Agenda concept in the user model) as an important, yet unexplored source for this category of situations. Users often introduce their social events in a calendar application (Google, Outlook). The framework proposes a mechanism to the user to de ne the reachability preference for a social event. This can be considered as an extension of the calendar template, where typically one speci es the name of the event, the participants and the time interval. Our framework proposes a mechanism that prompts the user before the event to specify his/her reachability preference. The only requirement for this to work is that the user lls regularly in his/her calendar application the upcoming social event. Currently an implementation for mobile devices is deployed and we are gathering user feedback. The implementation also contains a user-friendly interface that allows to tag the social network category of a contact. The tagging happens after an interaction with the given contact through pop-ups with buttons (Friend, Professional, Family...). This allows the user to rapidly include all contacts into a category, which is an important requirement for the system to work. Screenshots of the current implementation are shown in Figure 2.
Our partial feedback resulting from experiences made on 5 persons during di erent time periods has allowed us to validate the fact that social network categories play an extremely important role in such preferences. Typically, in the case of a professional meeting almost every person using the application blocked phone calls from friends. Family seems to be the category that is always allowed to interrupt. Another observation is the fact that preferences change very dynamically. Initially we thought that preferences change only for general situations (like work or home) and this seems to be wrong. A preference changes even when the user moves from one place to another, like from the o ce to the co ee machine. Currently this can be detected only when the user adds into the calendar this event, which is not very frequent. We are currently investigating how to detect such minor situations. Another major issue is the privacy concern, since the context of the users is logged. An e cient mechanism is needed to allow the algorithm to be incremental and thus to erase systematically the gathered context logs. 4
We address in this paper the question of how to build a user pro le model that can help better control our reachability in the context of an increasing exposure to unexpected interactions (spam messages, incoming calls that are not important in a situation etc.). We present a dynamic, situation-aware user model framework that enables real-time situation detection and a better control on the user's reachability. We propose mechanisms (extraction, classi cation algorithms, user feedback and services) that allow the model to adapt to the user. Future work include a real evaluation for the current implementation and improvement of the data mining algorithm. Also, gateways to social networking sites are considered to better automatize the categorization of contacts, leveraging that speci c sites consider di erent spheres of the social network and that these systems are open for this kind of information retrieval.