Section: Scientific Foundations

Keywords : recommender system, personalization, collaborative filtering, hypermedia, user profile, KDD, CBR, case-based reasoning, experience management, reuse of past experiences, social navigation, search engine, search access, indexing, user behaviour.

Supporting Information Retrieval with adaptive recommender systems

We think that information retrieval support tools as recommender systems are very useful in very large information systems. The objective of a recommender system is to help system users to make their choices in a field where they have little information for sorting and evaluating the possible alternatives [83] , [79] , [71] .

A recommender system can be divided into three basic entities (cf Figure  3 ): the group of recommendations producer agents, the module of recommendation computation and the group of recommendations consumers.

Figure 3. Architecture of a Recommender System

A major challenge in the field of recommender systems design is the following: How to produce adaptive recommendations of high quality minimizing the effort of producers and the consumers?

Two main complementary approaches are proposed in the literature:

1. approaches based on the content and the machine learning of user profiles and

2. approaches known as a collaborative filtering based on data mining techniques.

The user profile is a structure of data that describes user's centers of interest in the space of the objects which can be recommended. The user profile is a structure built in the first approach or specified by the user in the second approach.

The user profile is used either to filter available objects (content-based filtering), or to recommend to a user something that satisfied previous users with a similar profile (collaborative filtering) [79] .

In the Axis project, we continue the development of a hybrid approach for recommendations based on the analysis of visited content and on collaborative filtering; The past behaviours of a user group are used to calculate the recommendations (collaborative filtering). Like this this approach is able to support some usage evolutions without a complete re-design.Also the usage analysis of such recommender systems may be very useful to support designers in an possible re-design or improvment of their IS.

Approaches based on data mining are mainly statistical approaches where the sequence of events in the history is not taken into account for the calculation of recommendations. There are some early examples in the field of navigation assistance on the Web: the FootPrints system [86] and the system of Yan et al. [87] .

The implementation challenges of our approach relate to the following aspects:

• providing techniques of identification and extraction of relevant behaviours (i.e. the learning behaviours or case behaviours) starting from raw data of past behaviours,

• defining methods and measures of similarities between behaviours,

• defining inference techniques of adaptive recommendations starting from the identified relevant past behaviours (or starting from the reminded cases).

We study the class of recommender systems, based on the re-use of a user group's past experiences, using case based reasoning techniques (CBR).

Let us remind what is Case-Based Reasoning (CBR). It is a problem solving paradigm based on the reuse by analogy of past experiences, called ``cases''. In order to be found, a case is generally indexed according to certain relevant and discriminating characteristics, called ``indices''; these indices determine in which situation (or context) a case can be re-used.

Case-Based Reasoning  [76] usually breaks up into four principal phases:

1. a ``retrieve'' phase for cases having similarities (i.e. similar indices) with the current problem,

2. a ``re-use'' phase where a solution to the current problem is built, based on cases identified in the previous phase,

3. a ``revise'' phase where the solution may be refined with an evaluation process,

4. a ``retain'' phase that updates the elements of the reasoning by taking into account the experiment which has been just carried out and which could thus be used for future reasoning.

Difficult problems in CBR are related to: definition and representation of a case, organization of the database containing the cases, various used indexing methods and definition of ``good'' similarities measurements for the case search, link between the steps research and adaptation (the best retrieved case being the most easily adaptable case), definition of an adaptation strategy starting with the found case(s), training of new indices, etc.

We focus on two types of recommender systems:

• systems where the calculation of recommendations is based on the re-use of an users group's experiences in searching for information in a hypertext information system like the Web or on an Internet/Intranet site. These systems aim at an adaptive assistance to the search for information activity ;

• systems where the calculation of recommendations is based on the re-use of past experiences of experts, in order to provide an assistance to the design process.

We explore all three problems previously described by using case-based reasoning (CBR) techniques and more generally KDD techniques.

We pursue the evaluation of our results in CBR, in particular the indexing model by behavioral situation, the object-oriented framework CBR*Tools and toolbox Broadway*Tools via our current contracts (cf. section  7.1 ). Moreover, we pursue the study of sessions indexing techniques and plan to use some sequential pattern extraction and clustering algorithms for the on-line and off-line analysis of the Web users usage.