SUPERCALCOLO
BOLLETTINO DEL CILEA N.93GIUGNO 2004 21
www.supercomputing.it:
a portal that interviews its users
Silvano Mussi
CILEA, Segrate Abstract
The paper presents a new feature the CILEA supercomputing portal has been provided with. The feature aims at enhancing the portal user-profiling skills. It adds to other Artificial Intelligence features the portal is already endowed with. The feature consists in the capability of interviewing users about their work activities in order to collect fine grained profile data. The need of this feature is motivated by the following problem. Very often it happens that some initially acquired (from the user) information is not sufficient to infer the user interests with low uncertainty. So the portal has to ask the user further questions, questions whose answers may require a certain mental effort from the user. These questions cannot therefore be asked all and indistinctly at the same time. In fact, given that each question entails a trade-off between expected benefit and expected disturb effect, the problem of sequential asking, driven by the criterion of the most opportune question, arises. The new portal feature represents a solution to such a problem. The framework of the Value of Information Theory based on Shannon’s Entropy underlies the adopted solution.
Keywords: Supercalcolo, web-portal, personalization, user profile, supercomputing
Introduction
The present work represents a continuation of a work activity aiming at providing the CILEA supercomputing portal (in the following, for short: portal) with the capability of “understanding” user interests in order to recommend right news to right users. In [2] the reasons why the portal has to be adaptable to heterogeneous user typologies are illustrated. In [3] and [4] the author faces the problem of building an adaptive portal, i.e. a portal that, equipped of probabilistic reasoning [6], infers multiple user interests from few information and recommends the most relevant news. In order to exhibit an effective adaptive behaviour the portal has to gather fine grained information from the user. If we consider explicit information only, the standard way to acquire it is that of asking the user. The work presented in this paper focuses on the capability of effectively eliciting information from the user, and represents an application, in the context of the portal, of the methodology presented in [5]. Both the previous work and the present one represent an application of a background work [1] aiming at integrating the expert systems technology with the Website
technology. The ideas underlying the proposal have been implemented by the author in an off-line prototype.
The portal acquires user profile data by
sequentially asking the user
After having acquired (from the user) some “easy to provide” profile data (e.g. the software packages used by the user), the portal examines the state of uncertainty (measured in terms of entropy) of some strategic hypotheses about possible user interests. If the information so far collected is not sufficient to infer the user interests with low uncertainty, the portal consider the possibility of disturbing the user by asking additional information that is not, in general, as easy to provide as the first one. For example, answering to a question like “Do you use the computational technique X?” might not be as easy as answering to the question “Do you use the software package X?”. In fact the user might have developed an approach that uses some variants of the technique X and contains some original aspects, and so forth.
The portal considers a question at a time. For each question (no yet asked), the portal considers expected advantages and expected
SUPERCALCOLO
22 BOLLETTINO DEL CILEA N.93GIUGNO 2004
disadvantages: if its expected benefit (i.e. expected entropy reduction) is greater than its expected disturb effect, then the question is added to the set of questions that are worth being asked. Finally, the question that turns out to be the most promising is asked. The portal acquires the answer and then, on the basis of it, reconsiders the whole situation.
The portal seizes favorable occasions for
eliciting further profile information
The expected disturb effect is a cumulative decreasing temporal function. So after a certain period of time it reaches the null value. As a consequence, it might happen that a question, which was not worth being asked in previous time, becomes the preferred one in subsequent time. Moreover, if a favourable circumstance occurs, the portal seizes the opportunity for resuming the profile data elicitation session (exhibiting this way a sort of tenacious and yet non-intrusive behaviour). Given that, the following question arises: what is considered “favourable circumstance” by the portal? If it happens that the topic of a current piece of news is the same as the one of a question not yet asked, the portal estimates if it is worth asking the question. If it is, then a favourable circumstance for asking the question occurs.
Acknowledgments
The author thanks the following colleagues (listed in alphabetic order):
Enrico Cavalli who integrated the prototype into the portal running on the server computer; Paolo Ramieri who played the role of CFD expert both providing CFD knowledge and user-side feed-back during the testing phase;
the systemistic staff (Enrico Cavalli, Federico Ferrario, Marco Pirola, Giorgio Quattrone, Diego Travaini) who provided systemistic assistance about the various software environments involved by the prototype.
References
[1] S. Mussi, "Sequential decision-theoretic models and expert systems”, Expert Systems, Vol. 19, No. 2, 2002
[2] S. Mussi, "Conceptual portal-features for heterogeneous user communities", Bollettino del CILEA, No. 84, Ottobre 2002
[3] S. Mussi, "Providing Websites with capabilities of one-to-one marketing", Expert Systems, Vol. 20, No. 1, 2003
[4] S. Mussi, “Automatic Personalized News-Recommendation: a new service of the CILEA Supercomputing Portal”, Bollettino del CILEA, No. 88, Giugno 2003
[5] S. Mussi, "Putting Value Of Information Theory into practice: a methodology for building Sequential Decision Support Systems”, Expert Systems, Vol. 21, No. 2, 2004
[6] J. Pearl, "Probabilistic Reasoning in Intelligent Systems", San Mateo, CA: Morgan Kaufman, 1988
