personalized-recommendation-tutorial-description

Informations

• Publié par: Urzoe
• Nombre de lectures: 20
• Langue: English

Extrait

AI Techniques for Personalized Recommendation
Proposal for a full-day AAAI 2004 tutorial
John Riedl and Anthony Jameson and Joseph Konstan
Recommender Application Space1 Brief Description of the Tutorial
Overview of several dimensions along which recom-Personalized recommendation of products, documents, and
mender applications vary (e.g., purpose of recommen-collaborators has become an important way of meeting user
dations)needs in commerce, information provision, and community
Initial discussion of privacy issuesservices, whether on the web, through mobile interfaces, or
through traditional desktop interfaces. This tutorial rst re-
Applications and Interfacesviews the types of personalized recommendation that are be-
Discussion of a representative set of types of applicationing used commercially and in research systems. It then sys-
and interface from three areas:tematically presents and compares the underlying AI tech-
e-commerceniques, including recent variants and extensions of collabora-
document recommendationtive ltering, demographic and case-based approaches, and
off-web interfacesdecision-theoretic methods. The properties of the various
techniques are compared within a general framework, so that
Case Studies of Applications and Interfacesparticipants learn how to match recommendation techniques
Several in-depth discussions of particular applicationsto applications and how to combine complementary tech-
and interfaces, in the form of:niques.
live demonstrationsThe full-day format makes it possible to include a session
screen shots of important systems that are no longerin which participants actively work together on a concrete
availableproblem, as well as in-depth discussion of application con-
texts, case studies, and key social issues. In addition to illustrating the concepts introduced in the
The tutorial presupposes a general knowledge of AI. Some previous section, the case studies raise further speci c issues
previous familiarity with issues of personalized recommenda- and encourage interactive discussion.
tion is desirable but not essential.
Overview of the Major Approaches to Personalized
Recommendation2 Detailed Outline
Initial presentation of a general integrative frameworkThis tutorial will be an updated version of a successful full-
for all of the techniques to be discussed in this and theday that was held at AAAI 2002 in Edmonton. We
following sectionsalso gave a half-day version of the tutorial at IJCAI 2003 in
The ve major types of recommendation technique areAcapulco, since the IJCAI tutorial program could not accom-
distinguished on an abstract level, and a preview is givenmodate full-day tutorials. Although the half-day version was
of the ways in which they complement each other byalso well received, we very much want to return to the full-
having different sets of typical strengths and limitations.day format for AAAI 2004, for reasons to be detailed below.
This fact in turn points to the value of hybrid methodsTable 1 shows the proposed schedule for the tutorial. The
that combine the strengths of different approaches.following subsections comment on the various parts of the
tutorial shown in the table.
2.2 Part 2: Techniques Based on Collaborative
2.1 Part 1: Introduction, Applications, and Filtering:
Interfaces It is explained that collaborative ltering will be discussed in
Introduction more detail than each of the other approaches, because many
Basic concepts and motivation of the general issues are introduced for the rst time in this
Tutorial goals and preview section and because more experience has been acquired in
Brief history of personalized recommendation this area with the practical issues that arise with deployed sys-
Background of presenters and participants tems.Table 1. Proposed tutorial schedule.
(This schedule presupposes that the sessions and coffee breaks begin at the same times of day as at AAAI 2002.)
Morning Afternoon
9:00−9:20 Introduction Lunch break
9:20−9:35 Recommender Application 2:00−2:45 Non−CF Techniques
Space 2:45−3:15 Non CF Case Studies
9:35−10:15 Applications and Interfaces 3:15−3:45 Hybrid Techniques
10:15−10:45 Case Studies of Applications Coffee break and Interfaces
4:15−5:00 Designing Recommender Coffee Break Applications
11:15−11:30 Overview of Recommendation 5:00−5:45 Advanced Topics Techniques
5:45−6:00 Conclusions, Questions 11:30−12:30 Collaborative Filtering
Techniques
12:30−1:00 Filtering Case
Studies
2.4 Part 4: Applying and Extending What HasCollaborative Filtering Techniques
Been Learned So FarBasic K-nearest-neighbor algorithms
Item-item collaborative ltering Designing Recommender Applications
Dimensionality reduction algorithms Instructions on how to choose a particular application
Explaining recommendations based on collaborative l- and work out several high-level aspects of its design
tering Discussion in groups of about 5 participants, with occa-
sional advice from the presenters
Collaborative Filtering Case Studies Brief reports from the groups on their ideas and the prob-
lems they encounteredThe case studies serve a function similar to that of the
case studies for applications and interfaces, but the em- Advanced Topicsphasis is on the recommendation techniques.
Presentation and discussion of selected recent develop-
ments in personalized recommendation, such as:2.3 Part 3: Techniques Not Based Solely on special issues raised by recommender systems for
Collaborative Filtering groups, such as the issue of nonmanipulable prefer-
ence aggregationNon-CF Recommendation Techniques extended uses of data obtained from recommender
Content-based methods, including those that make use systems
of case-based reasoning or text classi cation methods issues in mobile/wireless recommender systems
Demographically based methods, which base recom- meta-level recommender systems
mendations on personal characteristics of the customer anthropomorphism in systems
Utility-based methods, which predict the value of items agent participation in recommender
to each individual user on the basis of a model of the
Conclusions, Questionsuser’s preferences, which may have been elicited explic-
Discussion of any further issues raised by the partici-itly
pantsKnowledge-based methods, which incorporate some sort
of domain knowledge to match users’ requirements with
the properties of items 3 Necessary Background and Potential Target
Audience
Case Studies of Non-CF Recommendation Techniques
While the tutorial presupposes a general knowledge of AI,In-depth discussions of particular systems that embody it does not presuppose any particular knowledge of person-some of the techniques introduced in the previous sec- alized recommendation. Participants at the AAAI 2002 andtion IJCAI 2003 tutorials ranged from persons seeking an intro-
ductory overview to well-known researchers with importantHybrid Recommendation Techniques publications in the eld of personalized recommendation. All
General overview of types of hybrid techniques of these participants can bene t because of the integrative
Examples of applications of nature of the tutorial, which serves the objective Present anovel synthesis combining distinct lines of AI work . This of papers and posters from this broad area. But an unnec-
perspective results in the inclusion of a much broader range essary degree of fragmentation into subcommunities can be
of material than any single participant is likely to be familiar observed, each of which focuses on one type of AI technique,
with. such as: classical collaborative ltering, case-based reason-
ing, decision-theoretic methods, or techniques for preferenceWhereas previous tutorials have usually focused on one
elicitation. Moreover, many researchers in this area have lit-particular type of learning or inference technique, this tu-
tle experience with the practical issues that arise with the usetorial covers techniques ranging from all major extensions
of personalized recommendation – issues which place someof collaborative ltering to case-based reasoning, decision-
strong constraints on the use of AI techniques.theoretic approaches, and various types of hybrid model.
This tutorial was developed with the explicit goal of over-These techniques are discussed in relation to the wide vari-
coming this fragmentation. The very positive feedback fromety of research and commercial applications that employ AI-
the participants at AAAI 2002 indicates that the goal wasbased recommendation. The algorithms, applications, and in-
terfaces presented are illustrated through a set of case studies achieved.
drawn from both research and commercial systems.
The target audience comprises (a) attendees from industry 6 Resumes of the Presenters
who seek an authoritative update on recommendation tech- One of the strengths of this tutorial, particularly for the AAAInology; (b) AI researchers who see personalized recommen- audience, is the fact that it brings together broad experiencedation as an attractive application area; and (c) researchers with AI techniques for user-adaptive systems (Jameson) andalready working on one type of recommendation technology extensive research and industrial experience in the speci cwho would like to learn more about alternative approaches. personalization subarea of recommender systems based on
After taking this tutorial, attendees will (1) be familiar with a collaborative ltering (Konstan and Riedl). The result is a tu-
broad set of