This paper describes how to provide background information adapted to the model of a group of people present in an Active Environment. 1 A.S.D. BodyEnergy, Mola di Bari, Italy.
Today most public places are provided with large-screen, digital displays or other
output devices typical of the particular environment: examples are cardio-fitness
machines in a fitness center, displays of a food dispensers, bus/train/plane notice-boards,
etc.. In opposite to on-line information seeking, such displays promote the experience
of “encountering” the information while carrying on another activity [
In this paper we propose an approach to group modelling that aims at providing background information adapted to presumed information needs of people present in a communal space and to context features such as the particular activity zone in which information is displayed, the time of the day and so on. The system adopts an approach in computing the profile of the user group that considers the fact that people to which information is addressed may be totally unknown, or may be known in full or in part, for example if the profiles of all or some of the users may be transferred to the environment. In order to test the system we selected as active environment a Fitness Center1. This type of environment is interesting for the main purpose of this research since: i) people subscribe a contract with the center and, contextually, it is possible to ask them to fill a questionnaire about their interests; ii) users are often provided with magnetic badges that allows identifying their entrance in the environment; iii) users are heterogeneous and have different interests, furthermore for certain period of time their presence is quite stable with some turn-over periods; iv) the overall environment can be divided in different activity zones in which it is plausible that people have different information needs (i.e. reception, fitness room, locker rooms, etc.); v)it is possible to make a statistical forecast of how many and which categories of users are present in an activity zone in a given time slice and therefore to combine this information with the profiles accessible by the system.
In the rest of the paper we describe briefly the system architecture and the group
modelling strategy and , in the last Section, we discuss results.
The approach presented in this paper is an evolution of GAIN (Group Adapted
Interaction for News) [
MusicFx [
The study involved 170 subjects (typical representation of the target users of our application during a day. Subjects were requested to fill a questionnaire that was divided in three sections aiming at: i) collecting some demographic data about the subjects (gender, age, category); ii) understanding the frequency of attendance (at what time, for how long and how many days during the week subjects were attending the place) and in which activity zone subjects are supposed to stay in each time slice according to the habitual activity performed in the place; iii) understanding which were the possible topics of interest by asking subjects to score them in a list using a 1-5 Likert scale (from 1 “ I really hate it” to 5 “I really like it”) for each activity zone. From this data we derived some information about the habits of every user in term of average number of minutes spent in each activity zone during a day, and about their distribution in every time slice. Figure 2 shows, in particular, the distribution of subjects’ interests when being in the Fitness room in two different time slices: around 10.00 a.m. and 18.00 p.m..
These time slices were selected as being quite different: in the morning the fitness room is attended prevalently from women while in the early evening from young male students.
Then, the definition of the group profile is made Figure 2. Comparison of interests in two time slices according to the formula we propose in (1) where different weights may be assigned to the known and unknown groups, according to the relative importance one wants to give to one group with respect to the other. In (1), we denote as: - PSURE, the weight (from 0 to 1) given to the preferences of known group. - PPROBABLE2 , the weight (from 0 to 1) given to the preferences of the unknown group. - K, the number of topics; - UMij the score for topic j in the activity zone A from user i; - b the base of the votes; can be 1, 5, 10, 100, etc.. - N , the number of known users; - M the number of profiles that constitute the statistical dataset (initially M was equal to the number of profiles collected in the preliminary study); - fi , the frequency of the attendance for every user of the selected activity zone, calculated as the number of days attended by user i divided by number of working days; - ti , the average number of minutes in which the user i is in the activity zone in the considered time slice; - , the frequency in the statistical dataset;
Then, Cj, indicating the confidence value for a topic j to be shown to the group in the activity zone A, is computed as follows: 3 (1) with N>0 and M>0 and b <>0.
This formula is a variation of the Additive strategy [
In the context in which GAIN is applied, the group model can be updated in different
situations: in a non-interactive context, in an “collective” interactive context using
a public touch screen display, in a personal interactive context through a personal
device. In all cases, we believe it would be impracticable to update the model by
asking people to explicitly rate the relevance of background news, especially if they are
in that place for a different purpose [
In order to validate our approach we tested the impact of the adaptation strategy first through a subjective evaluation study and then using a simulation of the system behaviour. This first study involved 80 people that usually attend the Fitness Center. One group of 40 people received a non-adapted random selection of news while for the other group news where selected according to the statistical profile. Both groups received news on a non-interactive display. We performed a t-test on the two datasets and results showed that the group that received news adapted to the statistical profile found in average more interesting (p-value=0,016), appropriate (p-value=0,004) and adequate (p-value=0,01) the proposed news than the group that received non-adapted ones. The second experiment, conducted using a simulation realized with a multiagent system, was aiming at understanding how to tune Psure and Pprobable values according to the different situation that may occur. Results show that if profiles of known users are statistically different from the statistical one then the list of the news to propose to the group will vary considerably. This becomes even more evident when Psure is high. In this case a plausible criteria could be to set Psure according to the ratio between the number of known and unknown users. If known users are similar to the statistical profile, then the mark does not change very much and, obviously, a high Psure may enforce the position of some categories in the classified results. The Psure value can be used to carry weight the context features Results obtained so far seem to confirm that the mixing statistical information about the group with those about known users allows to handle efficiently news adaptation in active environments.