CHAPTER 1. INTRODUCTION Given the variety of articles sel ected fr auser(bcth catent-b ased and ccllabara- tiverecamendatins both exploratory adexplaitativechaices, selections pertinent to differet topics of interest), a significant problemis to winnow down and present a set of recommendations in an interface The issue here inannutshell, is to investigate way s of represeting these tradeoffs. and to design interfaces that b leae the user in cantral of the carpasiticn of the recommendation set, and at the sare tire gather sufficient feedb ack to enable the Systentolean abetter user rod 1.3 Summary of contrib ution s Theresearch issues desai edhav e been addressed throug h the design aud leer tation of a seies of protcty pes, which have been tested in a vaiety of ex paimats The resul ting anginal research contributions are suarez ed in the fallowing four paints. An architecture linking populations of adaptive software agents, which takes advantage of the overlaps of interests among users of a recommender system to increase efficiency and scalability This ardntecturedy nanically links ev diving ad adapting papulaticms of ag ats to a changing uer commity. It all ows far scaling up the mmber of usas while maur taining afixedievel of resaurce usage proiding agraceful degradation in the qual ity ofrecamrendatians. This is achieved by encourag ing the evclutianof agents to serve clusters of user interest, ad pooling the feedbac relating to such clustas. Its dis- tributed nature ensures robustness and allows far users and agents to be added a rero ed d narically
8 CHAPTER 1. INTRODUCTION Given the variety of articles selected for a user (both content-based and collaborative recommendations, both exploratory and exploitative choices, selections pertinent to di erent topics of interest), a signi cant problem is to winnow down and present a set of recommendations in an interface. The issue here, in a nutshell, is to investigate ways of representing these tradeo s, and to design interfaces that both leave the user in control of the composition of the recommendation set, and at the same time gather sucient feedback to enable the system to learn a better user model. 1.3 Summary of contributions The research issues described have been addressed through the design and implementation of a series of prototypes, which have been tested in a variety of experiments. The resulting original research contributions are summarized in the following four points. An architecture linking populations of adaptive software agents, which takes advantage of the overlaps of interests among users of a recommender system to increase eciency and scalability This architecture dynamically links evolving and adapting populations of agents to a changing user community. It allows for scaling up the number of users while maintaining a xed level of resource usage, providing a graceful degradation in the quality of recommendations. This is achieved by encouraging the evolution of agents to serve clusters of user interest, and pooling the feedback relating to such clusters. Its distributed nature ensures robustness and allows for users and agents to be added or removed dynamically
13. SUMMARY OF CONTRIBUTIONS A novel recommendation mechanism combining a content-based and a collaborative method A methad is introduced which coi ines both content-based and collab arative ted niques, such that many of the disadvantages of using either approach al me are car celled out. This rethod is appl ied to the text recamerdatian task in the context of the arctecture desaibed aboe. It cam al so be considered as a wa far collab arative Sy stens to be applied to previouly prcbleratic dy naric drains, such as recommendation of Web pages ar news articles A new affordance allowing users to control the breadth or narrowness of their set of recommendations A large-scale simulated setting was cartructed riraily to investig ate particul ar plantations far explaratory and exploitative strategies the resul ts of the ex paints lead to a fuller understanding af how such strateg ies can be iplemted and how they impact speed of learning user rodels, respmsiv eness to change im user interests and the canpositin of recamedaticn sets. The result of this re- search is an implemted user inteface conponent allowing users to select "broad ar"narow'recorendatims, carespanding to exploratory ar exploitative strate- gies, respectivEly A new interaction design allowing users to vary the proportions between topics of interest, where only implicit feedback is required An implemented iterface based n a novel carpanent composed of sliding pah Els, allows users to card the prapatins between ultipletopics of interest, thu providing a finer-grained contrcl over the composition of recommendation sets. The
1.3. SUMMARY OF CONTRIBUTIONS 9 A novel recommendation mechanism combining a content-based and a collaborative method A method is introduced which combines both content-based and collaborative techniques, such that many of the disadvantages of using either approach alone are cancelled out. This method is applied to the text recommendation task in the context of the architecture described above. It can also be considered as a way for collaborative systems to be applied to previously problematic dynamic domains, such as recommendation of Web pages or news articles. A new a ordance allowing users to control the breadth or narrowness of their set of recommendations A large-scale simulated setting was constructed primarily to investigate particular implementations for exploratory and exploitative strategies; the results of the experiments lead to a fuller understanding of how such strategies can be implemented and how they impact speed of learning user models, responsiveness to change in user interests and the composition of recommendation sets. The result of this research is an implemented user interface component allowing users to select \broad" or \narrow" recommendations, corresponding to exploratory or exploitative strategies, respectively. A new interaction design allowing users to vary the proportions between topics of interest, where only implicit feedback is required An implemented interface, based on a novel component composed of sliding panels, allows users to control the proportions between multiple topics of interest, thus providing a ner-grained control over the composition of recommendation sets. The
CHAPTER I. INTRODUCTION interface further enables learning of user models through implicit gathering of feed back, by monitoring and interpreting user drag-and-drop actions(thus imposing much nitive load upon users 1. 4 Ov ery iew of the rem ain der of this thesis Chapter 2 defines concepts and introduces notation which will be useful in the re- mainder of the thesis. In doing so it also discusses a number of assumptions made thus situatin s research relative to various academic communities debates and schools of thor Chapter 3 specifies a very simple content-based recommender system This serves to introduce the specific search and machine learning algorithms employed, and to highlight problems which are solved by later, more sophisticated systems. It corre- sponds to an early working prototype called"LIRA"(Learning Information Retrieval Agent), and results are shown from a user study conducted with it. The design of LIRA and its testing were originally published as [Balabanovic and Shoham, 1995 and more fully as [ Balabanovic et al., 1997 In Chapter 4, alternative schemes based on collaborative filtering are introduced This is a short chapter as a purely collaborative prototype was not constructed as part of this research A multiagent architecture as referred to in section 1.3 is introduced in Chapter 5 as implemented by the "Fab prototype. This is the most complete of the implemenT tations discussed, and was available for public use on the Web for several months well as for more controlled user tests The first two contributions listed in section l3 are presented in this chapter; this work was originally published as alaba novIc 1997] and [Balabanovic and Shoham, 1997 Chapter 6 goes on to discuss problems related to selecting a recommendation set
10 CHAPTER 1. INTRODUCTION interface further enables learning of user models through implicit gathering of feedback, by monitoring and interpreting user drag-and-drop actions (thus imposing much less cognitive load upon users). 1.4 Overview of the remainder of this thesis Chapter 2 de nes concepts and introduces notation which will be useful in the remainder of the thesis. In doing so it also discusses a number of assumptions made, thus situating this research relative to various academic communities, debates and schools of thought. Chapter 3 speci es a very simple content-based recommender system. This serves to introduce the speci c search and machine learning algorithms employed, and to highlight problems which are solved by later, more sophisticated systems. It corresponds to an early working prototype called \LIRA" (Learning Information Retrieval Agent), and results are shown from a user study conducted with it. The design of LIRA and its testing were originally published as [Balabanovic and Shoham, 1995] and more fully as [Balabanovic et al., 1997] In Chapter 4, alternative schemes based on collaborative ltering are introduced. This is a short chapter as a purely collaborative prototype was not constructed as part of this research. A multiagent architecture as referred to in section 1.3 is introduced in Chapter 5, as implemented by the \Fab" prototype. This is the most complete of the implementations discussed, and was available for public use on the Web for several months, as well as for more controlled user tests. The rst two contributions listed in section 1.3 are presented in this chapter; this work was originally published as [Balabanovic, 1997] and [Balabanovic and Shoham, 1997]. Chapter 6 goes on to discuss problems related to selecting a recommendation set
1. 4. OVERVIEW OF THE REMAINDER OF THIS THESIS ndpresentinng it totheuser. Accordingly, section 6.1 concens aniplenetaticmaud siml ated ex periments to investig ate the properties of the exploration/exploitation tradeoff(tobe publ ished as [B alab ano ic 1998a), aud section 6. 2 concerns thedesign ofauser interface to gather implicit feedbad ad givethe user cantral ofrecameH daticms rel ating to ltipletopics of interest(also described in[Bal abano ic 1998b Bal abano ic, 1998d) Finally, Chapter 7 wraps upwith conclusions andrev isits the contributions above. Related wark is discussed throug hout rather than in a separate chapter
1.4. OVERVIEW OF THE REMAINDER OF THIS THESIS 11 and presenting it to the user. Accordingly, section 6.1 concerns an implementation and simulated experiments to investigate the properties of the exploration/exploitation tradeo (to be published as [Balabanovic, 1998a]), and section 6.2 concerns the design of a user interface to gather implicit feedback and give the user control of recommendations relating to multiple topics of interest (also described in [Balabanovic, 1998b; Balabanovic, 1998c]). Finally, Chapter 7 wraps up with conclusions and revisits the contributions above. Related work is discussed throughout rather than in a separate chapter
CHAPTER I. INTRODUCTION
12 CHAPTER 1. INTRODUCTION