In this paper, we expose a set of initial ideas related to an innovative way of structuring and organizing personal information. Indeed, users have to deal with a huge amount of information either coming from social connections, collected on the Web or generated by them. This phenomenon leads to new research challenges. In particular, how to structure, organize, and classify this personal information in order to better manage the user's digital memory? In this position paper, we present the concepts of activities and intentions as means for the user to structure efficiently all his past information, but also help him in the future, for example by suggesting relevant events, anticipating his information needs or providing opportunities to satisfy latent desires.
media, massive content production or diversification of knowledge sources, users tend to be overwhelmed with a huge volume of personal information such as emails, photos, e-books, blogs, social feeds, or various documents. These data are either created by them (e.g. through lifestream aggregators such as FriendFeed1,
social exchanges, information gathered on the web, etc. and constitute what we call the user’s digital memory.
However, today this information is only captured, stored, but not very-well organized from users’ point of view and thus is not used as much as it could be. This phenomenon induces the following research challenges. First, how to keep track of important events? Which semantic structure would allow users to find the right information when needed and organize their digital memory properly? A second
challenge deals with the anticipation of information needs: we believe that a usercentric semantic organization of the digital memory may help the user in his current or future information needs.
Thus, we present some initial ideas towards a new way of indexing and structuring users’ digital memories. Section 2 gives an overview of existing models and solutions for managing personal information. Section 3 introduces the notion of activity as a key concept to structure personal memory. Section 4 gives some clues on how to go beyond this first layer, by enriching this semantic structure with an additional metalayer of information organization, based on the notion of intention. Section 5 illustrates how this intention-based personal information management model can be instantiated for improving content filtering and opportunistic recommendations.
field has already been studied in the domain of personal information management,
and several paradigms of document organization have been identified. Temporal
paradigms organize documents according to a time line. This is the way how life
streams3 [
relationship a person engages in (books, music, photos, video, office documents, email, phone calls, meeting, web pages, etc.) and structure them according to two axes: time and life (personal vs. professional). However, according to Gemmel, “the collection is so large that the user cannot remember much of the contents, and will never use them.” Some solutions use a spatial representation, such as in Data
engines such as Google Desktop4 are an alternative to structured information, but in the case of digital memory, they do not rely on an index with the right granularity from the user’s point of view. Other approaches propose manual ways of structuring information. For example, Pearltrees5 proposes to users a way to keep content they find everyday on the web and to let them structure their information through trees.
information needs. Thus PackHunter [
However existing work are limited to an organization through a structure (e.g. timeline, hierarchical) with limited semantics which does not correspond effectively to the way users behave. So, there is a need to better structure this digital memory to make it useful and usable to the user. In this paper, we propose a solution using episodic memory [12] with two different layers: activities of the user and his intentions. We detail these concepts in the following sections.
3
In the human memory process, two main steps are fundamental: the acquisition
(retention) and recall. Tulving in [12] showed that episodic memory, which receives
and stores information about temporally-dated episodes and spatio-temporal relations
among them, is a faithful record of a person’s experience. Recalling a piece of
information is easier when the user can remind himself in time and space. Besides,
according to a recent study [
Capture user activity User timeline past Information indexation, information filtering Present future Information retrieval, anticipation of user information needs
current user activity. Capturing user activity is a research area in itself, where different related work [13] could be used. The “past” side of the timeline enables to index content and people and keep track of user memory. Past activities are reference marks (i.e. episodes) for people to find information and content, and a support for social information sharing even after their end.
More formally, we define an activity as a personal activity (digital or not) or as a user’s perception of a given social activity or event. Based on this definition examples of activity can be: reading a book and making notes and comments, or meeting someone in a conference and exchanging information, collecting multimedia content related to a user activity. An activity is composed of the following main properties: - A set of content that the user has generated, consumed or bookmarked in the context of the activity. A consumed content can be any type of multimedia content or web bookmarks. A user generated content can be an important piece of information written about the user activity (document, comments and annotations, notes) or any interaction captured during the activity (phone call,
- A semantic context is inferred from the set of content. It is a key enabler for the awareness of the activity community, and for further information classification. - A social context of the activity is the list of people that are sharing this activity (implicitly people around the user), or people following this activity (explicitly defined by the user or gathered from interaction traces related to the activity semantic context). - A spatio-temporal context of the activity. Time and place are the two dimensions that can be used to identify typical user contexts such as “at home”, “at work”, “on the move” or simply to position the activity in space and time for a better user recall. - A status. An activity can have three distinct statuses: ended, ongoing and in mind. The ended status means that the activity belongs to the past and that it can be used as a piece of memory. An ongoing activity constitutes a recipient for new incoming information. An in mind activity is not yet started; this is used to describe latent activities that may be recommended in the future to the user.
The role of the activity is twofold: (1) a working space environment where all pieces of information (documents, emails, bookmarks, etc.) and pertinent contacts are gathered within a same structure, becoming a relevant index (on people and content) for structuring the user digital memory, and (2) a representation of the social environment of an activity, helping people to share information in a controlled way and to get information from their social networks around this activity.
The management of personal information through the notion of activity provides
already a first organization layer. However, it does not consider interdependencies
between activities. So, we propose to extend this semantic structure with the concept
of information container as a semantic entity that encapsulates a set of coherent
activities that are correlated according to the different activity dimensions. Ultimately,
the observation of correlated activities may denote user’s intentions in time and space,
that describe what the user wishes to achieve at a high and pragmatic level [
The “past activities” of the user (Fig. 1) are structured through an additional layer, an information container (Fig. 2). The latter is composed of a set of activities and one or several properties describing the nature of the correlations between activities: − A content link reflects the shared semantic context between all the activities; − A social link contains the common contacts or social context (family, colleagues, etc.) between the activities; − A logical link indicates how an activity relates to others. Possible links are causality (an activity is the follow-up of another one), temporality (an activity is the repetition of another one), etc.
Based on the analysis of these semantic links an intentional link can be inferred between the activities present in a given information container. An intention can be seen as the high level “glue” between several activities and describes the set of activities as a whole unit as in [11]. Contrary to previous works such as [14], we do not express an intention by a formal plan; nevertheless at a high level, it may be described thanks to an action verb, a complement and an intensity reflecting its certainty or feasibility.
In addition to its structuring role of past activities, the information container can be
seen as an active recipient, in charge of helping the user towards the “future” side of
the timeline (Fig. 1). Indeed, intentions act as a guideline that leads the user
involvement through various activities. Thus, the knowledge of existing intentions
can be used to recommend information associated to activities belonging to the
container or which are completely new for the user. Additional exploitations of
intentions can be envisaged through some forms of collaborative mechanisms for
different purposes, for example: 1) to enrich / suggest activities to a user based on the
detection of a common activity pattern with other users – this may help the user to
find faster what he needs; and 2) to build a dynamic social network around people
having a common intention, in order e.g. to help them to realize it jointly [
agent that enriches the information it contains with coherent new elements coming from specified information streams (email, IM, RSS feeds, notifications etc.).
− Construction of information containers: how to correlate activities to build those information containers? When a new activity appears, to which information containers should it belong to? Is it just a clustering problem? How are we able to modify the information containers if we detect an anomaly? − Identification of intentions: detection of a precise user intention may be difficult. A possible solution is to use a learning model, where the user at the beginning explicitly describes the intention associated to an information container. After a while, the model could suggest the user relevant action verbs and extract knowledge from social and/or content links as complements. Another possibility would be to use a collaborative model which compares information containers of one user to the ones of other users to suggest possible intention labels. − Monitoring of intentions: how to infer the progress with respect to an intention or an information container? − Usage and acceptance – how to capture or confirm user activities (what is the part of automation and manual declaration) and present information containers to users? 5
In this section, we express through a scenario how the semantic structure described above can be used, in particular as a way to go beyond classical recommendation systems. Fig. 4 summarizes the different user’s activities that occur during the scenario. This scenario shows how a system can monitor in real-time different user’s activities, such as watching a documentary, browsing the web, meeting friends, etc. and the nature of the resulting intentions over the time. “top news about Cambodia” X “go on holidays to Cambodia” “write a report on Khmer art” Detection of possible intentions : “go to Cambodia” (20%) “news about Cambodia” (50%) “write a report on Khmer art” (30%) Container creation with a semantic link:
Cambodia
X Intention confirmed + context “near bookshop”
+ context “meet friend”
Opportunistic recommendation of activity: “go to bookshop” recommendati on of activity: Proposition of new services (hotel, plane) “know lore about Angkor” “be informed about Cambodia”
INTENTIONS New event generates possible new intentions
ACTIVITIES aBbroouwtsKehwmeebr Loforkidfgoer a art and hotels
Watch Browse wneebar Angkor documentary about Khmer Caambbooudtia art
Buy book about Angkor
Collect ifnrfoom/afdrvieisneds Holidays in Cambodia about Asia: Organize travel roadbook, pictures, etc
− Activity indexing: from the user timeline several activities are detected and then indexed by the system based on their contexts (e.g. for the activity “watch a documentary” the semantic context is a documentary reference and its status is equal to ended). − Building of information containers: in the scenario the construction of the information container is quite easy as most of the activities share at least the same content link related to Cambodia (except the “search of a new fridge”). By correlating more precisely the existing activities with past activities from other users (based on a collaborative approach) a logical link can also be inferred from the same information container (e.g. travel booking). − Intention detection: within the Cambodia information container several user’s intentions may be inferred based on the underlying information container links. For each intention the system tries to formalize its meaning (e.g. verb + complement form). In addition to the previous treatment a certainty degree is computed reflecting the current intention relevance according to several parameters (context, activities, etc). While new activities appear, the potential intentions are refined or simply removed from their information container. Thus, in Fig. 4, at the beginning three intentions were inferred, and at the end only one seems to be relevant: “go on holidays to Cambodia”. Nevertheless first inferences are already useful for proposing relevant content or services – especially in an opportunistic way, where the user may not have thought about himself (e.g. meet a friend). Another interesting property of an information container is that even if an intention is ended (e.g. the holidays are now finished) it is still open to new activities; thus new intentions can emerged (e.g. know more about Angkor). 6
organizing personal information. We believe that the concept of intention provides a relevant conceptual framework to anticipate user information needs, and opens the way to new service opportunities for context-aware multimedia content access and delivery. However we still need to understand if semantic and social contexts are appropriate indicators of relationships between activities to deduce user intentions.
captured activities. This new way of managing personal information may have a real social impact, e.g. by providing opportunistic interaction with people driven by intentions. To go a step further in the social exploitation, we envisage the use of collaborative algorithms for better inferring intentions through the co-relation of activities.
of people sharing the same kind of intentions, which will ease social interactions, and help them collectively find the right path to fulfil it (joint realisation of an intention). A further perspective of this work could be the creation of communities of knowledge, based on people promoting their information container, and sharing with the community the solution they found. We could capitalize on this community of knowledge to identify similar patterns of activities to fulfil typical intentions, and propose appropriate compositions of services that can be seen as an intention-based service mash-up.