Work | MLIS Portfolio | Lit & Art | Outdoors & Travel | Personal | Home  

Home >> MLIS >> 289 >> Adaptive Websites >> Full text

Your Colleagues Found These Additional Resources Useful: Adaptive Websites Applications and Libraries

Nov. 8, 2002


Introduction - Adaptive Websites and Personalization

Most library websites don't fare well against the better e-commerce sites such as Amazon.com™, Barnes & Noble (www.bn.com), and CDNow™ (www.cdnow.com). From complicated search interfaces to spare presentations (compared to many commercial sites), the library sites are too often aimed at those who know exactly what they want and lack the serendipitous possibilities of browsing the shelves or even thumbing through the old card catalog. Users researching topics, particularly the general user or those in the first stages of research, may benefit more from the wealth of connected information on Amazon's interface. Unlike the typical library website, Amazon can suggest related books to its customers, including those that others buying the same book also bought, along with lists, other media, and more - and the suggestions change over time as the customer purchases or browses for other items. One of the keys to Amazon's success is its reliance on an adaptive website that learns from both the explicit choices its customers makes in their personal profiles and their buying habits.

Amazon's interface isn't the only flavor of adaptive design, of course. Applications of many varieties can customize interfaces by creating models of user preferences. They may fill out forms, like Gator's eWallet (www.gator.com) and Alexa (www.alexa.com), or sort email, like Eudora's filters. Yahoo (www.yahoo.com) is just one of many companies who have developed portals that allow users create a customized interface, including preferences for news feeds and other content from within and outside the site.

This paper examines the development of adaptive [1] Web and applications in the commercial real, with suggestions for possible uses in a library website. Can libraries take advantage of adaptive technology to retain and attract patrons through personalized web applications? To what extent does personalization conflict with the strongly held value of privacy in libraries?

Defining the Adaptive Website

In a static website, every viewer will see the same interface, exclusive of differences due to browsers or screen size. Despite the underlying adaptive feature of the Web -Yahoo vice president Henry Sohn points out, "clicking on a hyperlink is personalization - you decide where you want to go" (qtd. in Lasica, 2001, Personalization defined), users will not be able to change the content within a site. At most, they may be able to customize browser settings to alter the text size, link appearance and other presentation characteristics. In addition, each time they visit the site, they will get the same content, at least until site is updated. Brusilovsky (2002) proposes the answer is to "develop systems with the ability to adapt their behavior to the goals, tasks, interests, and other features of individual and groups of users" (p. 30). He notes, "Adaptive web systems and hypermedia belong to a class of user-adaptive software systems" (p. 30) and defines them as collections of information resources that allow users to navigate between items and search for the useful resources. Within this context, Blom (2000, Intro.) defines adaptation as "process that changes the functionality, interface, information content, or distinctiveness of a system to increase its personal relevance to an individual". He describes personalization as user-initiated (explicit - e.g. based on preferences on a form or options dialog box), system-initiated (implicit - e.g. based on a user's actions), or a combination of the two (e.g. the system waits for the user to authorize changes). In addition, McKay (1990, p. 153) notes the changes should carry from one session to another. The goals of adaptation include providing only information that a user is - or might be - interested in (Nichani & Rajamanickam, 2000, What is personalization anyway?).

The system-initiated changes may encompass collaborative filter-based recommendation systems like those found on Amazon or CDNow (Nichani & Rajamanickam, 2000, What is personalization anyway?). The user's actions are compared with those of other users and when similarities are found, the system can offer suggestions.

Kühme, Dietrich and Malinowski (1993) propose a matrix of adaptability, in which the adaptation process is broken down into four stages: initiative, proposal, decision, and execution. Either the system or user can control each of these stages, so that the resulting adaptability is a result of a combination of system and user actions, rather than originating solely from one or the other. Blom (2002) concurs with this collaborative view, noting that personalization is "important to view as a dimension rather than a dichotomy" (Intro.). Brusilovsky (1996), in defining adaptive hypermedia systems as "systems which can provide automatic adaptation on the basis of the user model" (p. 109), argues that a slightly different framework is more useful for hypermedia, such as Web applications. He describes adaptability in three stages: "collecting data about the user; processing the data to build or update the user model; and applying the user model to provide the adaptation" (p. 109). In his model, the final stage is by definition performed by the system

The Development of Adaptive User Interfaces

One of long-standing goals of computer scientists and the artificial intelligence (AI) community has been to create machines that can adapt to user's needs rather than force them to adapt to hardware and software limitations. The efforts have taken place largely within the area of interface design and intelligent interfaces that emulate human intelligence in human-computer interactions. Researchers at Xerox PARC (expanding on earlier research at Stanford) and other labs helped bring relatively user-friendly computing to the masses by the mid-1980s with development of the Macintosh, and later, Windows (Myers, 1998). Many of the ideas of the computer science pioneers like Doug Englebart and Ted Nelson ultimately blossomed into ubiquitous graphical user interfaces, mice and other pointing devices, hypertext (based in turn on Vannevar Bush's memex idea (1945), and the Web.

Prior to the development of GUIs, the major effort in creating intelligent interfaces focused on natural language dialog - a metaphor based on the idea of a "human-like agent" behind the scenes, with whom the user conversed in a query and response pattern. (Encarnação, 1997, Intelligent user interfaces: intro.). As GUIs became common, more effort was put into developing applications which took advantage of direct manipulation. Even though natural language and other intelligent agent approaches have had fewer successes, each area has strengths that complement the other. Miller, Sullivan and Tyler (qtd. in  Rauch, 1997) noted in 1991 that the goals of intelligent user interfaces include:

These goals remain important today. With the spread of computers, the range of demands on them increases, in terms of both the variety of users and the types of applications (Hirschman, 1999).

Much of early adaptive interface design relied on prior user modeling, or stereotypes, to bring human-like intelligence to the computer. Yet with limited success in human emulation, researchers found a collaborative approach more practical (Fischer, 1999, sect. 4). In this approach, the system will more often make suggestions instead of managing changes directly. A fairly recent example is Microsoft Office, in which interactive help characters offer contexualized tips - usually a more useful feature than the non-contextualized random tip-of-the-day brought up by opening an Office program. Fischer (1999, sect. 1) comments that a challenge of user modeling for complex applications such as Office is providing a rich variety of options and functionalities - few of which may be wanted by any single user, but any of which may be important to some user - without the unneeded functions getting in the way, yet remaining available if needed.

The traditional static hypermedia, whether a web page or an electronic encyclopedia, offers the same view to all users, and also the same view for successive uses. By the early 1990s, researchers recognized the problems with this one-size-fits-all approach and began looking at combining ideas from hypertext with user modeling. Brusilovsky (2001, p. 88-89) identifies 1996 as the year when a major shift occurred in the field that had become known as adaptive hypermedia. Several research teams began important projects, a number of workshops and conferences were held, many more papers were published, and the first PhD dissertations in the field were written. He ascribes the change to the use of the Web, which provided a real-world platform with a "widely diverse audience" for research, and the existence of a body of research work large enough to provide a basis for further research.

Brusilovsky's 1996 (summarized in Brusilovsky, 2001) review of hypermedia identified six kinds of systems - "educational hypermedia, online information systems, online help systems, information retrieval (IR) hypermedia, institutional hypermedia, and systems for managing personalized views in information space" (p. 89). He notes that nearly all the development since 1996 has taken place in the first two areas, with some significant progress in IR hypermedia (which now includes systems for personalized information views); almost no interesting work had been done between 1996 and 2000 in help systems or institutional hypermedia.  In his view, online information systems now encompass:

With the influence of the web, the latter two have diverged far enough to be considered separate categories. Likewise, IR hypermedia has broadened to include:

Greenspun (1999, chap. 9, Case studies conclusions) identifies development of the Netscape Magic Cookie protocol in 1994 as the key to many of the subsequent adaptive web developments. Cookies allowed servers to collect a variety of user information that becomes useful in user modeling. Amazon's personalization is based largely on tracking these cookies.

Brusilovsky (2001) lists the user characteristics that have most often served for user modeling: the "user's goals/tasks, knowledge, background, hyperspace experience, and preferences" (p. 96). In addition, new applications may apply user interests (both short- and long-term, serving as a basis for additional relevant resources) and the user's individual traits (primarily in education hypermedia - he notes that few successes have been obtained, in part due the difficulty in determining which traits are important). Finally, a fair amount of development has taken place in environmental adaptation such as presenting different interfaces depending on the device, physical location (through GPS), movement, or other characteristics.

Adaptive Web-Based Applications

This section examines two current web-based applications employing adaptation in commercial settings. The Washington Post (www.washingtonpost.com), like most other major news outlets, has explored filtering to better meet readers' interests. CDNow, the largest online music retailer makes use of an effective recommender system to suggest new music to customers. Providing individualized experiences has been viewed by both organizations as key to continued success.

The News You Need: The Personalized Washington Post

M.I.T.'s Nicholas Negroponte (1996) coined the "Daily Me" (p. 153) to describe a personalized news media that would provide only the information that a user wanted, whether world events, politics, or horseracing news. Instead of a handful of regional editions of the New York Times, every subscriber would see his or her filtered version. Negroponte has been pushing the concept since the mid-1970s (Hapgood, 1995, ¶ 5). Reseachers in the M.I.T. Media Lab began developing systems that were both customizable and personalized, with the idea that an agent would be able to search both traditional and non-traditional sources for news of interest to the user (Bender, 2002, p. 22). Beginning with News Peek in 1981, which included implicit user models and filters. Network Plus followed in the mid 1980s, and FishWrap in 1993. The latter utilized Doppelgänger, a component that created dynamic user models through inferring preferences from a variety of databases profiling individuals' interests, beliefs, behaviors, and more. FishWrap also included PLUM, a component which makes use of freely available geographic information to contextualize news stories by geography, demographics and weather history (Bender, 2002, p. 23). 

Negroponte's concept had been largely unrealized outside of the Media Lab until the development of the Web. In the mid 1990s, PointCast launched the first personalized news and information service, with a number of "My" portals such as MyYahoo soon following (Lasica, 2002, ¶ 3). Most of the resulting sites were, however, limited to explicit customization requiring users to fill out lengthy forms.

In 1998, usability expert Jakob Nielsen (1998) wrote that personalization was overrated; matching technology was insufficiently developed to "safely predict what stories will be of most interest to me" (¶ 10), although he expected that it could happen within ten years. By most measures, Negroponte's vision is beginning to come true on a large scale with new personalized news services from the Los Angeles Times, the New York Times, and a number of other major publishers.

The Washington Post launched its MyWashingtonPost in on June 5, 2002. Since the papers' readers - even its print readers- have a wide variety of interests, the basic site may require a fair amount of clicking to locate desired stories. With personalization, a basic filtering can take place with the minimum of information that MyWashingtonPost requires users to provide when they register, including zip code (for localized weather, traffic, etc.) and birthday (for horoscopes). Lasica (2002, The Washington Post) notes that the Post gradually accumulates additional user information to tailor the content. In addition, the user can choose to get information from over 500 categories by making changes to his or her profile. Finally, the user can make changes to the presentation on
the fly by moving around the elements and adding links to outside resources.

Fig. 1. Basic MyWashingtonPost interface
Fig. 1. Basic MyWashingtonPost interface

Post vice president of personalization efforts, Tim Ruder, proposes that the renewed efforts for personalization among news organizations and other businesses have been spurred by a realization that unless they can engage their audience, their audience will go elsewhere (Lasica, 2002, The Washington Post). Weinberger (2002, p. 45), in discussing the meaning of Web space, notes that just because a mega-corporation like AOL-Time-Warner creates a complex site or set of sites, absolutely nothing prevents visitors from going to another site they like better - the Los Angeles Times is only a click away from the Washington Post. Although the technology has improved to make things easier for users, Ruder (Lasica, 2002, The Washington Post) claims that the key is developing stronger and deeper relationships with users. So far, this effort to increase user loyalty is working. Ruder notes the paper has met its registration targets and has garnered much positive feedback from readers who have said the changes have helped them discover previously invisible information buried deep in the site.

Ultimately, the Post and other companies entering the personalization game hope that the individual attention will translate into revenue, particularly from specialized users. Lasica (2002, The New York Times) cites the New York Times, which is exploring personalized services (like Microsoft's .Net initiative) to deliver a rich, mobile information system to professionals and business users.

Some commentators, have criticized the Daily Me concept for taking away the power of editors, but Lasica (2001, The horror) and others see that power shift as a prime reason to celebrate it, positing that it puts the power of choice with the individual because the Internet isn't a mass medium, but rather a medium for the masses that encourages both one-to-one and one-to-many communication.

Some critics have also expressed concern about the Daily Me as a force for fragmenting society.  Cass R. Sunstein, law professor at the University of Chicago, argues in his 2001 book republic.com that news personalization leads to information ghettos and uninformed citizenry. He observed the trend toward extreme views among groups of like-minded people on the web and elsewhere; the ability to filter can exacerbate this trend (Kaplan, 2001). Others dismiss Sunstein's concerns. Lasica (2002, intro.) believes that those who personalize will be likely cast their net wider and deeper than those who don't; we are so awash in information that even the narrowest filters won't shield us from major events and news. Even though the Post might prefer its readers to get all news from its pages, sources of news on the Web are multiplying - particularly the less traditional ones that are unlikely to feed the Post or other mainstream outlets.

Fallows (2002, sect. 2) notes that Sunstein retracted in part his statement that the Internet is dangerous to democracy following a discussion in pages of the Boston Review; proposing instead that the Internet in fact allows people to encounter many new ideas. Kaplan (2001) points out that it is relatively easy to shield ourselves from marginal ideas and people in the real world. Mainstream media does a fine job of filtering already. The Daily Me unleashes in each of us "our basic desire to share," writes Landers (2002, p. 29), "which sometime translates into a sharing of information, social and political ideas, or goods and services. We are more deeply engaged in learning, more in tune with our priorities, and ever expanding our scope."

It's Your Music: Recommender Systems at CDNow

CDNow's strength, like Amazon's is based on a successful recommender system. Such systems help address the problem of too many choices and too little time to explore them. (Rashid, et al, 2002). Since its inception in 1994, CDNow has encouraged customers to provide music preferences by broad category so that the company could target its marketing. It followed Amazon's lead in making personalization a centerpiece of its business strategy, with features including:

Hof, Green and Himelstein (1998, Changing focus) reported the wish list feature proved immediately successful in increasing the number of visited pages. Rather than include an explicit community review and ratings feature like Amazon's, CDNow has chosen to highlight staff reviews, but encourages bulletin board postings for opinions and discussion.

Since its debut, CDNow has added a variety of personalization features including group modeling that builds recommendations from ratings (on music the customer either purchases or already owns). The system will also make recommendations based on past purchases. CDNow uses a combination of user- and item-based collaborative filtering based on the user's purchases and explicit preference to make recommendations that lead customers to explore - and buy - music that they might never have otherwise considered or even known about.

CDNow's newest personalization effort, MyCDNow, launched in Sept. 1998, with a customizable home page with suggestions based on past purchases, stated preferences, and ratings on CDs and artists. When customers open the MyCDNow interface, they can choose to view their wish list; favorite artists; recommendations; or edit their email, music and movie preferences (the company began offering DVDs and videotapes in Dec. 1999). A tabbed interface allows users to switch back to the standard views at any time.

Fig. 2. MyCDNow interface
Fig. 2. MyCDNow interface

Schafer, Konstan and Riedl (2001, p. 2) note that recommender systems in e-commerce are used to make suggestions for purchase based on a variety of data, including top sellers, customer demographics, and purchase predictions based on analysis of past buying patterns. They describe the forms of recommendation as including product suggestions, personalized product information, providing community opinion and critiques for products. The recommendation lists may be created by human editors or by the system. This personalization makes possible an individual store for each customer. The recommender systems encourage browsers to buy (by suggesting specific products of interest), cross-selling (similar to items placed next to a checkout line in a grocery store), and building loyalty by matching customer interests (most successfully when the system learns through repeated visits, thereby increasing the match rate). In addition, customers can send their own recommendations to friends via an email form.

Even active customers, observe Sarwar, Karypis, Konstan and Riedl (2001, Challenge of user-based), may have purchased only a tiny percentage of the very large total offerings; the recommendations from user-based collaborative filtering may not be a close fit for any particular user. In addition, new items may not show up in recommendations because they haven't been rated before (the "first-rater" problem) (Melville, Mooney & Nagarajan., 2002, p. 187). This problem isn't absent from CDNow, but because they are the largest online music retailer, they have richer data to exploit for collaborative filtering.

CDNow has a number of advantages for personalization over many other online businesses. Their targeted consumers are more likely to be interested in contributing information about themselves and their favorite music than typical customers at on online gardening site might be. By using a variety of recommender systems, CDNow can make better suggestions than if they used only one. This strategy has proved successful in keeping their top place among online music retailers.

Surfing to MyLibrary

The web interface for Stanford University Libraries (SUL) has been essentially unchanged for several years. It is utilitarian and spare, with few graphics or fancy presentation. From the home page, the user can access a variety of resources, including the catalog and databases, pages of collections and resources, user information (including circulation records), computing help, and a number of library enterprises. Socrates, the library catalog, is available in a public, as well as in a Stanford-only version. Other than the access to circulation records and course reserve catalogs, no part of the site is personalized. Each visit is the first ever.

Fig. 3. SUL home page
Fig. 3. SUL home page

Rich Holeton, director of residential computing, commented in a presentation on SUL's web interface [2] that a growing number of undergraduates go first to Amazon to research their paper topics. Can the techniques used in e-commerce and online news be applied to academic library websites to make them the first destination for their users?

Amazon's success makes it the obvious model for redesigning the present interface and some library vendors have worked toward this goal. The recommendations, multiple types of resources, and other features presented on Amazon's site could also be helpful in the library context. A number of vendors provide Amazon-like interfaces. Sirsi Corporation, for example, has created iBistro and iLink, which are respectively personalized portals for public and academic libraries. Both products incorporate a number of the features listed above, including recommendations (Sirsi Corp., 2001).  Although Jakob Nielsen (qtd. in Livingston, 2001, Don't slavishly follow Amazon's model) warns that Amazon's model may not fit every need, the differences between a bookstore and a library are small enough that Amazon's model might work well with a few adjustments. The information-rich presentation might contain:

Fig. 4. Socrates home page
Fig. 4. Socrates home page

The Amazon approach is likely to be most successful in a public library, where a general readership could benefit from the recommender systems that would encourage patrons to discover new books, magazines, other materials from the library's collection. In a research library, the primary benefit would perhaps be more limited to undergraduates; graduate students, professors and other researchers will generally not need system-generated recommendations since they will already know their field well and will have other ways of keeping abreast of new publications. The library catalog for them is most useful in verifying holdings rather than resource discovery. Nonetheless, even they might find some of the personalization features useful.

Privacy may be the biggest issue in library website personalization. While privacy is an issue in e-commerce, Amazon and other businesses do keep databases of customer records and provide personalized services by utilizing both the information in their databases and cookies on their customers' computers. Libraries also collect certain patron information, but most attempt to balance security and privacy needs as laid out in the ALA Code of Ethics (1995) and Access to electronic information, services and networks (1996). Kobsa (2002) argues that users must be able to interact with personalized systems in an anonymous or pseudonomynous manner. Although the US currently lacks strong privacy protection, the EU and some other countries prohibit identifiable user profiles. Sunderham's discussion (2002, p. 351) suggests client-side personalization is the way to avoid abrogating privacy, and in a research library, the trend toward wireless access and individual computers make this approach more possible. Rodríguez-Mulà, García-Molina and Paepcke, (1998, sect. 1), Greenspun (1999, Quiet server-side personalization), and others note that the use of cookies makes client-side data gathering easy (as long as patrons permit cookies on their own machines).

Client-side personalization is less practical at kiosk or public cluster computers. Caches are often cleared either when a user logs off or on a regular basis [3]. Schafer, Konstan and Riedl (2001, p. 14) suggest session-based, ephemeral personalization would be a step beyond non-personalization. Although ephemeral personalization would not be able to tailor the interface with a patron's long-term needs and patterns, it would be able to offer some useful recommendations or adaptations during the session. The higher degree of adaptation found in longer-term persistent personalization would be practical only on users' own machines.

Personalization promises many benefits to library users, but only if carried out in a responsible manner so that privacy and confidentiality were maintained. The issues of trust and responsibility are important for businesses; they are doubly so for public and educational institutions. While users generally prefer personalization that doesn't require extra effort on their part (i.e. filling out long forms) (Price, 2001, intro.), the rising awareness of the possible misuse of personal information collected by firms such as DoubleClick has made users more wary of implicit personalization (Mabley, 2000, p. 1). In addition to the need for confidentiality, any information that is collected must be done for clear and specific purposes.

As privacy concerns grow, more research into methods of providing both privacy and personalization is taking place. Work such as that carried on by Canny (2002) (collaborative filtering through distributed computing) and Mobasher, Dai, Luo and Nakagawa (2002) (creating anonymous aggregate profiles) promise ways of giving library users the rich experience they can find at Amazon without loosing confidentiality. Until privacy can be guaranteed, full personalization for library websites should remain an unrealized goal. Nonetheless, libraries can take the lessons of adaptive website in e-commerce to make their own websites the first destination for patrons.

Footnotes

  1. The term "adaptive" is sometimes used interchangeably with "assistive," or the approach of creating interfaces that may be used by people of varying abilities. Disabilities access is a rich area for research, but lies outside the scope of the present paper. While the "adaptive Web" is only one facet of  "adaptive hypermedia," the discussion focuses on Web applications and the terms are used interchangeably within this context.
  2. I am indebted to Rich Holeton, Head of Residential Computing and his presentation on Oct. 30, 2002, Why I Don't Use Socrates Anymore, for a discussion of some of the listed items and ideas about redesigning the Stanford University Libraries catalog interface.
  3. At Stanford, no permanent cookie or other session files are maintained on the public cluster computers because disk images are loaded daily. Sessions are authenticated and logged through a secure, Kerberos login, but these logs contain user rather than activity information, to help correlate suspicious activity with a user during a specific time period. (Stanford Student, 2001, Academic computing).