Introduction to the Special Issue: Recommender Systems
Francesco Ricci and Hannes Werthner, Guest Editors
International Journal of Electronic Commerce,
Volume 11, Number 2, Winter 2006-07, pp. 5.
Recommender systems are intelligent applications that assist users in a decision-making process when they do not have sufficient personal experience to choose one item from an overwhelming set of alternative products or services. The scope of recommender systems has expanded gradually since their first introduction in the mid-1990s. Recommender systems were originally proposed as tools where “people provide recommendations as inputs, which the system then aggregates and directs to appropriate recipients” . This definition refers to the recommender system technology known as collaborative or social filtering, which has the merit of having started the research in the field and has inspired a number of hybrid approaches combining the idea of social filtering with other technologies [1, 4]. Research on recommender systems has historically overlapped with many computer science topics, but two main areas initially influenced that research: information retrieval and artificial intelligence. From information retrieval, recommendation technology research derives the vision that users searching for recommendations are engaged in an information search process and query a content repository in order to obtain a collection of matching items, typically in the form of a ranked list. From an artificial intelligence perspective, recommendation is viewed as a learning problem that exploits past knowledge about users, such as the search or buying behaviors of a community of users, in order to infer the current user’s interest in items not yet considered.
With this early imprinting, many of the first recommender systems were designed to support simple information-search-oriented human-computer interactions where two phases can be identified: user-model construction and recommendation generation. The user model, a structured description of the user’s needs and preferences, is typically acquired by exploiting either a collection of previous user-system interactions or information provided by the user during the recommendation session. The recommendation generation then reduces to the identification of a subset of products that “match” the user model, according to an ad hoc definition of the matching function. For instance, in collaborative filtering the user model comprises the ratings provided by a user for a set of products, and recommendations are computed by identifying a set of similar users according to their user profiles and then recommending products highly rated by similar users. This simple behavior, made up of a user-model definition and matching, deviates from a more natural human-to-human interaction, where the user and the recommender (adviser) interact by exchanging requests and replies until the user accepts a recommendation. Focusing on these aspects, conversational recommender systems have been proposed. In these systems, there is no clear separation between the user-model construction and the recommendation-generation stages. Conversational systems support a dialogue wherein the system has many alternative moves at each stage-it can ask the user for preferences, or request feedback on a product, or suggest products. Hence, research on recommender system has moved considerably toward a human-computer interaction perspective. The recommender is no longer regarded as an oracle that can predict the user’s tastes and suggest the “right” option; it is, rather, seen as an “adviser” that can leverage multiple factors to influence and even persuade the user’s decision process. This evolution is exemplified very well by the papers in this Special Issue, as we shall note below in describing them and their contributions to the research field.
First, however, we wish to emphasize that recommender systems are among the most typical outcomes of the Internet, Web, and e-commerce “revolution.” In fact, even if the problem of supporting a choice-decision process is quite old, it is only with the advent of the World Wide Web that we have had at our disposal, in large quantities, two basic ingredients of recommender systems: a large catalog of products, as in e-commerce Web sites, and a large population of consumers/users querying these Web sites and leaving the tracks of their on-line behavior. Since their introduction, recommender systems have been exploited for recommending books, CDs, movies, news, electronics, travel, financial services, and many other products and services . Recommender systems are becoming more and more popular even in rather simple e-commerce Web sites. The editors have been astonished by how this concept and technology, quite obscure in the travel and tourism market only five years ago, when we attended the 2001 ENTER conference on information technology, travel, and tourism in Montreal, are now being exploited by so many public and private Web sites in this market (leading in B2C with respect to volume) .
The papers in this Special Issue illustrate the wide range of issues under investigation today. The article by Felfernig, Friedrich, Jannach, and Zanker on knowledge-based recommender systems focuses on complex products such as digital cameras and financial services, and the requirements they impose on conversational recommender systems. In this scenario, the complexity of product definitions and assortments makes the identification of appropriate items a challenging task. Customers can differ significantly in their expertise and level of knowledge, so an intelligent recommender system must provide personalized dialogues supporting the customer in the product-selection process. The authors present successfully deployed commercial applications and their evaluations. And they illustrate a range of domain-independent, knowledge-based technologies exploited in their original recommendation environment, CWAdvisor. The authors show how model-based diagnosis, personalization, and intuitive knowledge-acquisition techniques support the effective implementation of customer-oriented sales dialogues.
McGinty and Smyth’s article illustrates another approach to conversational recommender systems-the acquisition and exploitation of product preferences by means of feedbacks and critiques by the user on products displayed by the system. This approach contrasts with more classical ones where the user’s preferences are asked, at the outset of the interaction, in order to construct a query and retrieve interesting products. By aggregating user feedbacks over a number of cycles, the recommender system can obtain a clearer picture of the type of product the user may be interested in. In these approaches the key problem is to keep the dialogue focused and short-that is, to reach a stage where the product is selected by the user after a small number of cycles. The authors illustrate how this “dialogue length” can be greatly reduced if the recommender system exploits an adaptive approach for selecting the products shown at each cycle. The system must offer more diverse recommendations when the user seems unable to provide useful new feedback. In contrast it must focus on more similar products when this feedback is acquired and does provide new information with respect to the previous cycle.
The third paper, by Ahn, presents a novel approach to product recommendation that uses the popularity characteristics of products. Popularity characteristics describe, for instance, how often a product has been purchased or whether it has received high ratings from consumers. This kind of information plays a significant role in the consumer purchasing process, but little attention has been paid to the use of popularity in recommendation research so far. The author describes a movie recommender system that integrates popularity and genre information about films to build a user profile that gives each genre and popularity type a personalized score on the base of the analysis of previously rated films. The recommender system can then propose new films belonging to the highly scored genre and popularity types of the user. The author compares this approach to the collaborative filtering method, showing that the exploitation of genre and popularity information can show positive results under data sparsity and cold-starting conditions.
The fourth paper, by Gretzel and Fesenmaier, contrasts with previous works where the main goal was to propose new methodologies and techniques for “computing” accurate recommendations. Here the authors investigate an issue that is receiving more and more interest-the role played by the recommendation process itself in convincing, and even persuading, the customer that the recommended products fit his or her needs and preferences. The authors investigate the potential influences of the relevance, transparency, and effort required by the preference elicitation and recommendation process on the perceived fit of the output recommendation. The fit of a recommendation is also correlated with the user’s perceptions of the elicitation process, measured by the perceived enjoyment and perceived value of the process. In this paper the authors point out some very important findings indicating that the relevance, transparency, and length of the preference-elicitation process serve as important cues for the experienced value, which in turn influences enjoyment of the process and the perceived fit of the recommendation with one’s preferences. The lesson learned from this analysis is that practitioners must carefully tune the perceived relevance and transparency of the supported recommendation process if they want to fully exploit their sophisticated recommendation algorithms.
The paper by Nikolaeva and Sriram investigates another important issue in recommender systems-measuring the value of the recommendation provided by the system and the characteristics of the user population as well as the impact of the product data distribution on this value. The authors build a model of internal information search for an unplanned purchase prompted by a recommender agent. They consider a recommender that supports the suggestion by saying that a similar user has bought the product the user is considering now. The model developed by Nikolaeva and Sriram describes how consumers update their beliefs about unplanned purchases upon receiving a recommendation. The increase in the expected utility of the purchase, after the recommendation, is called the marginal value of recommendation (MVR). The authors perform a Monte Carlo simulation to understand how these factors influence the effectiveness of the recommendation. They discover that the value of a recommendation depends on the preference structure of the recipient, the attributes of the product on which the recommendation is based, and the characteristics of the population of consumers. This analysis can be used to tune the recommender by operating on the characteristics that increase the value of the recommendation and are under the control of the marketer.
The last article of this Special Issue illustrates the relationships, mentioned earlier, between recommender system research and information retrieval. Many recommender system suggest products by searching in a data repository and matching user preferences with product characteristics. If products are described in an unstructured format, such as an HTML page, it is difficult to go beyond a simple syntactic match between the keywords in the query and the text. The paper by Lee, Chun, Shim, and Lee proposes an ontology to improve the quality of the match between a set of keywords and product descriptions in a B2B marketplace. Their method computes a score for each concept appearing in the product descriptions that takes into account the semantic relationships between concepts as defined in the ontology. A document that does not explicitly contain a keyword can be related to concepts that explicitly mention the searched keyword. Such an approach can be exploited not only to retrieve the best matching product description but also to generalize the search problem and to retrieve the most relevant classes of products. The authors shows that the application of a domain ontology can also improve retrieval performance when the user input is incomplete and not all the relevant keywords are used to search for the target product.
The Guest Editors thank three groups of people who have made this issue possible. First, the many reviewers (around 80) who not only volunteered to review the articles, but also provided valuable and extensive feedback within a tight schedule. Second, the authors who submitted their original work. Thirty-eight papers were submitted, and the quality of all of them was of a very high standard. They regret not having been in a position to accept more papers, mainly due to the space constraints of a journal issue. Third, the authors of the accepted papers for their speed and responsiveness in addressing the reviewers’ concerns; we were able to go through two revisions on all the articles and still keep the Special Issue on schedule. Finally, they also thank Professor Vladimir Zwass, editor-in-chief of the International Journal of Electronic Commerce, who agreed to take on this project and helped us to complete the task.
- Adomavicius, G., and Tuzhilin, A. Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE Transactions on Knowledge and Data Engineering, 17, 6 (2005), 734-749.
- Burke, R. Hybrid recommender systems: Survey and experiments. User Modeling and User-Adapted Interaction, 12, 4 (2002), 331-370.
- Resnick, P., and Varian, H.R. Recommender systems. Communications of the ACM, 40, 3 (1997), 56-58.
- Shardanand, U., and Maes. P. Social information filtering: Algorithms for automating word of mouth. In I.R. Katz, R.L. Mack, L. Marks, M.B. Rosson, and J. Nielsen (eds.), CHI 95: Human Factors in Computing Systems Conference Proceedings. Denver: ACM Press, 1995, pp. 210-217.
- Werthner, H., and Ricci, F. Electronic commerce and tourism. Communication of ACM, 47, 12 (2004), 101-105.
FRANCESCO RICCI (Francesco.Ricci@unibz.it) is an associate professor in the Faculty of Computer Science, Free University of Bolzano-Bozen, Italy. His research interests include recommender systems, intelligent interfaces, constraint satisfaction problems, machine learning, case-based reasoning, and software architectures. He is the author of scientific publications that have appeared in Communications of the ACM, IEEE PAMI, IEEE Expert, Applied Intelligence, International Journal of Intelligent Systems, and Information Technology and Tourism, and in numerous conference proceedings. In addition, he has co-chaired several national and international conferences, acts as a referee for IEEE PAMI, Machine Learning Journal, IEEE Transactions on Data and Knowledge Engineering, International Journal of Electronic Commerce, and Computer Journal, and is on the editorial board of Information Technology and Tourism.
HANNES WERTHNER (firstname.lastname@example.org) is professor of e-commerce at the Vienna University of Technology, and previously was professor of information systems at the University of Innsbruck and professor of computer science and e-commerce at the Vienna University of Economics and the University of Trento. He is the founder and president of the e-Commerce Competence Center (EC3) in Vienna, was a member of the strategic advisory board for the European research program IST in FP5 (ISTAG), and is editor-in-chief of Information Technology and Tourism. He is a visiting professor at the University of Surrey. He has a master’s degree and Ph.D. from the Technical University Vienna, and studied computer science at the Technical University of Vienna, was visiting professor at several universities, has published more than 100 papers and books, and was a fellow of the Austrian Schrödinger foundation. His research activities cover e-commerce, Internet-based information systems, decision support systems, simulation, and artificial intelligence. Key Words and Phrases: keywords