1. 1. THE RECOMMENDATION TASK research Readers with know ledge of the academic fields of information retrieval. collabora tive filtering, user mo deling and machine le arning will have recognized many of the issues raised so far. In the rem ainder of this int roduction the text recommendation tash will be defined more precisely and with respect to other more commonly studied t asks such as retrieval, routing, filtering, and cl assification. Following this definition the different research issues already alluded to will be revisited and explained at more det ailed level. What should be expected from the rest of this thesis? As befit s the practical nature of this t ask, a series of working prot otypes have been imple- mented and tested, both with real users and in simulat ion. By in crement ally defining these increasingly complex systems, the reader will gradually be introduced to the origin al contributions of the research. These range from methods for the collection phase (discovering documents of interest to a community of users) to designs for the interactions bet ween a recommen der system and a user. A summary of these contributions will form the final part of this introduction 1.1 hero, d atio task As is common for applications whose domain is the Web, this thesis dr aw s he avily on the traditions of the information retrieval(IR) community. However the t ask outlined diverges from the original goal of IR systems-n amely, retrieving documents to fu a short-term, specific in formation need as expressed in some query. In contrast, the goal of what h ave re cently come to be known as reconmmender systems Resnick and Varian, 1997] is to learn about a populat ion of u sers in or der to provide increasingly en dat ions, whe system whose domain is documents, we de fine the text recommendation task, following Jennings and Higuchi, 1993, as helping the user follow threads of interest, by regularly
1.1. THE RECOMMENDATION TASK 3 research. Readers with knowledge of the academic elds of information retrieval, collaborative ltering, user modeling and machine learning will have recognized many of the issues raised so far. In the remainder of this introduction, the text recommendation task will be de ned more precisely and with respect to other more commonly studied tasks such as retrieval, routing, ltering, and classi cation. Following this de nition, the di erent research issues already alluded to will be revisited and explained at a more detailed level. What should be expected from the rest of this thesis? As be ts the practical nature of this task, a series of working prototypes have been implemented and tested, both with real users and in simulation. By incrementally de ning these increasingly complex systems, the reader will gradually be introduced to the original contributions of the research. These range from methods for the collection phase (discovering documents of interest to a community of users) to designs for the interactions between a recommender system and a user. A summary of these contributions will form the nal part of this introduction. 1.1 The recommendation task As is common for applications whose domain is the Web, this thesis draws heavily on the traditions of the information retrieval (IR) community. However the task outlined diverges from the original goal of IR systems|namely, retrieving documents to ful ll a short-term, speci c information need as expressed in some query. In contrast, the goal of what have recently come to be known as recommender systems [Resnick and Varian, 1997] is to learn about a population of users in order to provide increasingly appropriate recommendations. When considering the speci c case of a recommender system whose domain is documents, we de ne the text recommendation task, following [Jennings and Higuchi, 1993], as helping the user fol low threads of interest, by regularly
CHAPTER. INTRODUCTION recommending small sets of documents (typically small means between 6-10). A tert recommen der system interacts with a user according to the following loop 1. Select a set of documents to p resent to the u ser, based on an induced user 2. Elicit feedback from the user. 3. Improve the u ser model given the feedb ack on the p resented documents The automatic layout of personalized new spap ers is an ong oing area of re seard (e. g, see Benderet al, 1991). This thesis, how ever, concent rates on the selection of artides for readers, ra ther than their final presentation and appearance as an electronic new spap er Several different IR tasks are defined by the Teart rEtrieval (TREC) conferen ces; it will prove u seful to appeal to these specfic defi nitionsin order to ground a di scu ssion of the differences between them and the recommendation task. In particular, TREO 5 Harman, 1996, the latest of this series of conferen ces at time of writing, will be The ad hoc retrieval task (sometimes called retrospective retrieval) requires the user to formulate a query for an IR system that accesses a particular corpus of doc uments. In re sp onse, the system ranks the documents according to rel evance to the query. The query rep resents a short-term information need, which will hopefully be satisfied by the top-ranked documents. In contrast, the recommen dation ta sk does not require a query to be formulated. Furthermore, discrete, u nordered sets of docu ments are delivered, rather than a ranking over an entire corpus. Finally, the iterativ nature of the re com mendation ta sk is better suited to long-term, ong oi ng informati needs. Ad hoc retrieval is the original task for IR systems, and is still thought of as canonical. It will be seen, how ever, that an IR"engine, a collection of alg ori thms and
4 CHAPTER 1. INTRODUCTION recommending smal l sets of documents (typically smal l means between 6{10). A text recommender system interacts with a user according to the following loop: 1. Select a set of documents to present to the user, based on an induced user model; 2. Elicit feedback from the user; 3. Improve the user model given the feedback on the presented documents. The automatic layout of personalized newspapers is an ongoing area of research (e.g., see [Bender et al., 1991]). This thesis, however, concentrates on the selection of articles for readers, rather than their nal presentation and appearance as an electronic newspaper. Several di erent IR tasks are de ned by the Text REtrieval (TREC) conferences; it will prove useful to appeal to these speci c de nitions in order to ground a discussion of the di erences between them and the recommendation task. In particular, TREC- 5 [Harman, 1996], the latest of this series of conferences at time of writing, will be assumed. The ad hoc retrieval task (sometimes called retrospective retrieval) requires the user to formulate a query for an IR system that accesses a particular corpus of documents. In response, the system ranks the documents according to relevance to the query. The query represents a short-term information need, which will hopefully be satis ed by the top-ranked documents. In contrast, the recommendation task does not require a query to be formulated. Furthermore, discrete, unordered sets of documents are delivered, rather than a ranking over an entire corpus. Finally, the iterative nature of the recommendation task is better suited to long-term, ongoing information needs. Ad hoc retrieval is the original task for IR systems, and is still thought of as canonical. It will be seen, however, that an IR \engine," a collection of algorithms and
11. THE RECOMMENDATJON TASK data stru ctures developed originally for ad hoc retrieval, proves u seful for a variety of other tasks More similar to the re commendation task are the routing and filtering tasks, where users are assumed to have long-term information needs belkin and Croft, 1992. In the routi ng task, the system is not provided with a query, but rather a set of sample docu ments that are relevant to a users information need. U sing these documents as training data, the system can then rank the remaining documents in the corpus. For the fil tering task, the system is instead required to classify the remai ning documents into categories of“ relevant”or“ irrelevant”. Variations of this task a re of ten stu died from a madinelearni ng perspective, as instan ces of supervised classification problems TYang, 1997--from a training set of sample documents with atta ched categ ory labels. a system learns to a ssign categ ory lab el s to new documents(for the fil tering task, the labels would be just“ rel evant”and“ irrelevant”) In contrast, a re com men der system mu st select its ow n samples of trai ning data to present to the user for feedback, thus engagi ng in active learning. Furthermore, the fact that the process proceeds in an incremental rather than batch fa shi on means that many of the tech niques used in routing or fil tering, which rely on batch process of the trai ning data. are no longer as appropriate. The difference between these batch task s and the recommendation task is anal og ous to that between sup ervi sed d assification and rei nforcement learning more recen t strand of re seard is collaborative filtering [Mal one et al., 1987 A coll ab ora tive system will recommend an item based on the judgments of other users, rather than on the content of the item itself. This is consistent wi th the lect recommenda tion task, and indeed Chapter 5 will present a way to combine collab orative filter ng with IR tech niques
1.1. THE RECOMMENDATION TASK 5 data structures developed originally for ad hoc retrieval, proves useful for a variety of other tasks. More similar to the recommendation task are the routing and ltering tasks, where users are assumed to have long-term information needs [Belkin and Croft, 1992]. In the routing task, the system is not provided with a query, but rather a set of sample documents that are relevant to a user's information need. Using these documents as training data, the system can then rank the remaining documents in the corpus. For the ltering task, the system is instead required to classify the remaining documents into categories of \relevant" or \irrelevant". Variations of this task are often studied from a machine learning perspective, as instances of supervised classi cation problems [Yang, 1997]|from a training set of sample documents with attached category labels, a system learns to assign category labels to new documents (for the ltering task, the labels would be just \relevant" and \irrelevant"). In contrast, a recommender system must select its own samples of training data to present to the user for feedback, thus engaging in active learning. Furthermore, the fact that the process proceeds in an incremental rather than batch fashion means that many of the techniques used in routing or ltering, which rely on batch processing of the training data, are no longer as appropriate. The di erence between these batch tasks and the recommendation task is analogous to that between supervised classi cation and reinforcement learning. A more recent strand of research is col laborative ltering [Malone et al., 1987]. A collaborative system will recommend an item based on the judgments of other users, rather than on the content of the item itself. This is consistent with the text recommendation task, and indeed Chapter 5 will presentaway to combine collaborative ltering with IR techniques
CHAPTER 1. INTRODUCTIO N fall, asite bosg sstesamsteg et al. 1995: Iiclen a 1995]po cea ae atietas mu lto whihsimnp Hiriestekaig piche. By e stith tesl estcteeta dtewchstata dtesesurethose kato. te emea ap. o stdio atetha alita web as 1.2 Research issues Itsdtu m ictceaapkedftedt aeaet atig a, eoaienesae fa∈ ae wleedaallthatis,on, it血eaee: oe a. atledlot a, e bgeax vitae is olete csesteto,, is estclei, estate b trs esah 1.2.1 How to exploit overlaps between users'interests Tese wlce @e teBGEa atileismeel oem ay usesdaeG melesse. PGhasteeaeaumled ole loei teste i atiks clot ieonehalle etedceitcind w astaaladle s@ that dup i ae <aching ad d soe, at itisaethniate-teeol leasigle "weep et p hasse hg this Omunit. Bute telynai ad aatie at edaeomeesste. itis, Eiatethats, etsaeascta@s es itdhesde edi: itol leati atoalozco,la.schdeo n(UOS动ae CO6Oi. Itmihttasi∈ttoO,BO∈ koeisakca ai,∈ae dati kselotw l, He ledietw ol le, essa, fsgedie ati- ksto Giiatefo a 1 lee et ifoeeits In geealay use ol haenulti,e koisdi, teet ait ail oela ig with ce useS An inp cta.t, ach isee iste Sin dsse stta taeav ataedthisstuctu eto
6 CHAPTER 1. INTRODUCTION Finally, assisted browsing systems [Armstrong et al., 1995; Lieberman, 1995] propose an alternative task formulation which simpli es the learning problem. By restricting themselves to the section of the Web just ahead of the user's current browser location, they recommend appropriate links to follow rather than arbitrary Web pages or documents. 1.2 Research issues Let us return to the example of the software agent creating a personalized newspaper for a reader, where so far all that is known is that the reader enjoyed an article about a new Bordeaux vintage. This context exposes the two main issues to be investigated by this research. 1.2.1 How to exploit overlaps between users' interests The user who enjoyed the Bordeaux article is merely one of many users of a recommender system. Perhaps there are a number of people who are interested in articles about wine. One challenge, therefore, is to nd ways to allocate resources such that duplicated searching and discovery activities are eliminated|there could be a single \wine expert," perhaps, serving this community. But given the dynamic and adaptive nature of a recommender system, it is desirable that such experts are a spontaneous result of the system evolving; it would be impractical to allocate or plan such decompositions in advance. Correspondingly, it might transpire that our Bordeaux lover is also an avid reader of articles about wildlife. Therefore it would be necessary for some of her articles to originate from a wildlife expert, if one exists. In general any user could have multiple topics of interest, arbitrarily overlapping with other users. An important research issue, then, is the design of systems that can take advantage of this structure to
1. 2. RESEARCH ISSUES perform dynamic load balancing, and so deliver efficiencies of scale as the number of users Incre ases In addition, it is interesting to investigate further conse quences of pooling feedback and dis covering clusters of interest. The con cept of coll aborative filtering has already been mentioned: a collabor ative strategy could, for inst ance, discover ot her users who had al so enjoyed the Bordeaux article, and then recommend other articles which they had liked. In contrast, a content-based strategy would recommend articles similar in content to the one about borde aux. By considering ot her users opinions of an article as well as its content, a hybrid approach can make more informed decisions. Therefore another aspect of this re search issue is how overlaps between users' interests can be ploited to improve recommendations, combining the st rengths of bot h coll aborative and content-based techniques 1.2.2 How to decide upon the composition of recommenda- tion set Techniques for choosing individual articles for recommen dation are well known, but deciding on the com position of a set of recommendations(or the selection of articles in a single edition of a per son alized newspaper) is an import ant issue that has re- ceived little at tention. A number of tradeoffs are involved, For inst an ce, as in the earlier example: the system comes across an article about a new str ain of flu. It has no formation about whether or not the user would like this article. since it has never re ceived feedb ack about any similar article. An exploratory strategy would re commend this article. In contrast, an exploitative strateg uld stick to articles similar to the known Bordeaux example. This is known as the explor ation/exploit ation tradeoff Berry and Fristedt, 1985
1.2. RESEARCH ISSUES 7 perform dynamic load balancing, and so deliver eciencies of scale as the number of users increases. In addition, it is interesting to investigate further consequences of pooling feedback and discovering clusters of interest. The concept of collaborative ltering has already been mentioned: a collaborative strategy could, for instance, discover other users who had also enjoyed the Bordeaux article, and then recommend other articles which they had liked. In contrast, a content-based strategy would recommend articles similar in content to the one about Bordeaux. By considering other users' opinions of an article as well as its content, a hybrid approach can make more informed decisions. Therefore, another aspect of this research issue is how overlaps between users' interests can be exploited to improve recommendations, combining the strengths of both collaborative and content-based techniques. 1.2.2 How to decide upon the composition of recommendation sets Techniques for choosing individual articles for recommendation are well known, but deciding on the composition of a set of recommendations (or the selection of articles in a single edition of a personalized newspaper) is an important issue that has received little attention. A number of tradeo s are involved. For instance, as in the earlier example: the system comes across an article about a new strain of u. It has no information about whether or not the user would like this article, since it has never received feedback about any similar article. An exploratory strategy would recommend this article. In contrast, an exploitative strategy would stick to articles similar to the known Bordeaux example. This is known as the exploration/exploitation tradeo [Berry and Fristedt, 1985]