This is a pre-print version of the paper! Please cite the published version as follows: Anke, J., & Sundaram, D. (2006). Personalization Techniques and Their Application. In M. Khosrow-Pour (Ed.), Encyclopedia of E-Commerce, E-Government, and Mobile Commerce (pp. 919-925). Hershey, PA: Information Science Reference. doi:10.4018/978-1-59140-799-7.ch148
“All users are created unequal”: Personalization techniques and their Application
Juergen Anke
[email protected] David Sundaram
[email protected] Department of Information Systems and Operations Management Private Bag 92 019 University of Auckland Auckland, NEW ZEALAND
Please address all correspondence to Dr. David Sundaram at: Department of Information Systems and Operations Management The University of Auckland Private Bag 92 019 Auckland, New Zealand eMail: Phone: Fax:
[email protected] +64-9-373-7599 ext. 85078 +64-9-373-7430
2
“All users are created unequal”: Personalization techniques and their Application Introduction Personalization is an approach to increase the usability of complex information systems and present the user a comprehensible interface which is tailored to his/her needs and interests. In this chapter we examine general techniques that are employed to achieve personalization of web sites. This is followed by presentation of real-world examples. It will be shown how different levels of personalization can be achieved by employing the discussed techniques. This leads finally to a summary of the current state in personalization technologies and the issues connected with them. The chapter closes with some ideas on further research & development and a conclusion. In general, the concept of personalization refers to the ability of tailoring standardized items to the needs of individual people. It is originally derived from the ideas of Pine’s book “Mass Customization” (Pine II, 1993). He claimed that companies should move from the paradigms of standardized products and homogeneous markets to customizable products which meet the requirements of many different customers (Schafer, Konstan, & Riedl, 2001). The principle of mass customization applies to a certain degree to most car manufacturers and some computer manufacturers, e.g. Dell. In the digital world of the World Wide Web the degree of customization can be much higher than in the physical world. Currently a number of online portals and e-commerce shops make use of personalization to provider a better user experience. Although web sites may be the most popular examples of personalization, it is not limited to the Web. Every information system that deals with large amounts of data and/or has a heterogeneous group of users can benefit from it. Examples include e-learning environments, electronic books, computer operated voice/telephony services, and tourist guides. Personalisation is also very useful for mobile devices like Personal Digital Assistants (PDAs) or mobile phones (cf. Mulvenna et al, 2000). Technologies like mobile internet access, WAP and future multimedia applications based on high-capacity wireless technologies require the designers of services for these devices to deal with limited input capabilities and small display sizes. For that reason every method that assists the user in navigating and finding information easily adds real value to applications for such devices.
Personalization Techniques The core idea of personalization is to customize presentation of information specific to the user to make user interfaces more intuitive, easier to understand and reduce information overload. The main areas of tailoring presentation to individual users are content and navigation. Content refers to the information being displayed and navigation refers to the structure of links which allow the user to move from one page to another. Personalized navigation can help the user to find easier what he is looking for or to discover new information. For 3
example, a system discussed by Belkin (2000) assists users in refining search queries by giving recommendations on related or similar terms. Adaptable vs. adaptive There are two approaches to achieve personalization: adaptable and adaptive methods. The former is a term for systems which can be customized by the user in an explicit manner, i.e. the user can change content, layout, appearance etc to his needs. This data is called “user profile” and all personalized presentation is based on data the user provided for configuration purposes. It is important to note, that the customized appearance does not change over time until the user decides to change his preferences. In contrast, adaptive methods change the presentation implicitly by using secondary data. This data can be obtained from a variety of sources, for example from the users actions, from the behaviour of other users on that site or based on the currently displayed content. Methods that use this data as input are discussed in detail below. The most distinctive characteristic of adaptive methods is that they are constantly monitoring the user’s activities to adjust the arrangement and selection of relevant information. Adaptive methods or machine-learning algorithms are huge steps towards automated customization. Current static interfaces suffer from the fact that the designer has to anticipate the needs, interests and previous knowledge of the users in advance. As these preferences change over time, customization that requires human interaction for collecting and identifying preferences leads quickly to outdated user-profiles. Table 1 shows how adaptive and adaptable methods can be applied to customize content and navigation. The examples given are intended to be generic; more concrete examples are examined in the case study below.
Adaptable
Content
Navigation
explicit selection and ordering of content items by the user
building link lists (favourites, bookmarks)
providing personal information to be listed in directories
setting default links for generic navigational structures/menus to omit intermediate step(s)
Adaptive
setting up stock portfolios present the user new items which are related to the current items (recommendations) filter content based on current actions (remove items which are dissimilar)
hiding unsuitable links based on the context annotate links to give metainformation about value of the linked content relating to the user’s navigation history (e.g. “no new information”, “insufficient previous knowledge” etc)
Table 1: Application of adaptable and adaptive methods to content and navigation 4
Degree of personalization Another important criterion for classification is the degree of personalization. System can have transient or persistent personalization or be non-personalized. With a transient personalization the customization remains temporary and is largely based on a combination of the user’s navigation and an item-to-item correlation. For example, if an item is selected, the system attaches similar items as recommendation to it whereby the content of the shopping cart is taken into consideration. Persistent personalization systems maintain a permanent user account for every user to preserve his settings and preferences across separate sessions. Although this raises privacy issues and is the most difficult to implement, it offers the greatest benefit. These systems can make use of user-to-user correlation algorithms and thus provide higher accuracy. The lowest degree of personalization can be found in non-personalized systems. However, this does not mean, that there are no dynamic elements in the presentation. The key point is that information is presented in the same way to every user. The majority of online shopping sites are still non-personalized. (Schafer et al., 2001) Another technology that belongs to the broad area of personalization is “recommender systems” (Mulvenna et al, 2000). Whereas straight personalization tailors just the presentation of information, recommender systems support the user in discovering new information. As recommendation relies on user preferences and interests it is often part of personalized systems. From another perspective one can say that recommender systems provide a selection of the most suitable content for the user. The application of recommender systems to e-commerce is discussed by Schafer et al (2001).
Application and Impact of Personalization: Amazon Amazon.com is one of the pioneers of e-commerce. Originally set up as bookstore it has grown to a general retailer for a wide range of products. It also provides an auction platform and a marketplace where customers can sell used goods. The marketplace is seamlessly integrated in the main product catalogue; therefore customers can decide whether they want to buy a particular product as a new or a used one. Goal
As the Amazon.com product catalogue contains more than two million products, users can easily get frustrated, if they do not find what they are looking for. Thus, one of the main goals is to tailor the product catalogue as far as possible to the needs and interest of the user. Aside from easy navigation the site offers seamlessly integrated recommendation system. It is intended to offer customers products which are either related to their interest or to the product which is currently displayed to exploit cross-selling potentials.
5
Employed personalization techniques
Amazon.com is a highly developed online shopping site and incorporates a combination of numerous adaptive and adaptable methods. The generated user profile is stored in a database on the server, i.e. the degree of personalization is persistent. The prevalent recommendation method is based on the purchases of other customers. It appears as “Customers who bought” list on each product detail page (Figure 1 , (1)). A second list contains up to five authors whose books were bought by customers who also bought the currently selected book. Both of these features are collaborative filtering methods of implicit collected data. Another feature called “Purchase Circles” is also a collaborative filtering mechanism that displays the “top 10” products for a particular region, institution or company.
6
(2) Individual bundle offers (1) Similar products
(3) Recommendation based on recently viewed items
(4) Aggregated rating of other customers (5) Own Rating
Figure 1: Amazon.com detail page for a book Other recommendations methods are based on the activities of the customer himself. For example, there is always one product recommendation in a sidebar which is apparently directly derived from the list of recently viewed items (Figure 1, (3)). Moreover, there is also a recommendation page for each of the main categories (books, software etc) which is completely based on the customer’s purchase history. Amazon.com also encourages customers to rate their purchased products (Figure 1, (5)). Ratings belong to the collaborative filtering methods. However, in contrast to the “Customer who bought” feature they require explicit interaction for the sake of personalization. The ratings can also be part of customer reviews and comments. 7
Interestingly it is even possible to “rate the ratings”, that is customers can indicate whether they found the rating helpful or not (“3 of 5 people found the following review helpful”). This mechanism is used to display most helpful comments first and let least helpful comments move to secondary pages. As mentioned before, one of the goals is to exploit cross-selling potentials. One of the recent additions to Amazon.com is the “Great Buy” feature. It offers the customer the current product together with another product in a bundle at a special price (Figure 1, (2)). The two products must have a complementary relationship to be valuable to the customer. It appears that this offer is based on products attributes rather than on customer behaviour or preferences. The algorithm that produces this output needs to find out a second product which is similar but not too similar to become a substitute. We assume that this is a relatively time-consuming computation which is done in a batch process rather than adhoc. An approach to accomplish this would be to determine similar categories and select products which either have been bought together often or are the best rated in the respective categories. Outcome and lessons learned
Amazon.com has clearly utilized a very broad range of different personalization methods. As this site has developed constantly in product variety, functionality and comfort it is nearly at the state of the art in this area. The goals stated above are nearly completely fulfilled. The site is easy to navigate, products are easy to find and the accuracy of the recommendation seems to be very high which animates the customer to buy further products. The “Great Buy” feature is certainly one of the best ways to take advantage of the crossselling potential, whereas the “Customer who bought” is more useful for discovery of new interesting items and navigation. The latter seems to be less accurate compared to the earlier days of the site. This might have something to do with the enormous number of customer profiles which do not provide enough distinctive attributes to form useful clusters. Ratings and reviews can be considered as a helpful feature; however there are a number of relatively “unqualified” comments. To improve the situation the “rate the rating” feature was introduced (far after the review function itself). While this highlights the more valuable reviews, there is still room for improvement.
Application and Impact of Personalization: Yahoo Yahoo was one of the first search portals on the web and one of the first websites that applied personalization in a larger scale [9]. In 1996 the “My Yahoo” service was introduced. It allows setting up a personal version of Yahoo for every user. Not only the content, but also the layout and the appearance of the page can be modified. Yahoo is a completely adaptable system; therefore all personalization is based on the data the user entered beforehand. Especially the ZIP code is central, as a lot of personalized features rely on it. The “intelligence” of the system lies in the ability to use this data in different situations to tailor the presentation specific to the user. 8
Goal
The goal of Yahoo is to bind its users by differentiating from other Web catalogues and search engines, and provide a fully customizable and integrated portal. As the customer structure of Yahoo is very heterogeneous it is a good idea to offer personalization and let users construct an individual start page. Yahoo’s service is free for its users; money is mainly earned with advertising and revenue provisions of shopping partners. Thus, the second goal is to make advertising as effective as possible. This can be achieved by selecting banner ads which are likely to be of interest for the user.
(4) Yahoo! Companion toolbar
(2) Weather module
(3) Delete or configure module
(1) Headline module
Figure 2: A customized version of the “My Yahoo!” start page Used methods of personalization
Yahoo offers an adaptable system which requires user to explicitly provide information for personalization. The user profile is kept on a server between different visits, thus Yahoo offers persistent personalization. “My Yahoo!” enables registered users to build their own Yahoo pages. The content is selected as so-called “modules”. Among the available modules are weather, news, sports results, stock quotes, horoscope, movie reviews, personal news filters and many more. Further configuration can be done within these modules. In the example shown in Figure 2 the headline module (1) is set up to show world news from British news agencies and German Formula 1 news. The weather module (2) displays the current weather situation of selected cities only. Modules can be edited or deleted (3) directly on the page. Some of 9
the modules offer an individual default setting which is based on the user profile. When the “Team results” module is added it already contains the results of teams in the user’s area. Not only the content but also the layout is customizable. The chosen modules can be distributed on multiple pages which in turn consist of two or three columns where modules can be ordered arbitrarily. There are also options like “Colour Sets” and “Themes” to change the appearance of “My Yahoo!”. Outcome and lessons learned
The goal of flexible customization to provide an integrated portal can be considered as fulfilled. Whether the effectiveness of advertising is increased by personalization can not be decided with certainty. The source [9] does not mention advertising at all. However, as it is vital for Yahoo to have effective advertising, it can be assumed that they incorporate a relatively advanced system for selecting banner advertisements on an adaptive basis. Apart from these sophisticated personalization features, it is also essential to design highquality default pages for people who do not want to customize at all. It turned out that only a small number of users actually customize Yahoo; most of them take what is offered. The reasons for that may be either that the personalization tools are too complicated or the users do not need complex personalization, as the default page is satisfactory for them. Addressing all types of users also includes not requiring users to enter any personal data or force them to use personalization features. Yahoo has decided not to use adaptive methods. They believe that these methods are still not good enough for such a complex site, as the user’s behavior is not sufficiently predictable. People must not be unsure how the systems work; otherwise it prevents them from experimenting, as they fear to break something. The people from Yahoo reckon that any kind of personalization should encourage the users to experiment. IBM
Another well documented large-scale personalization project is the ibm.com web site. Karat et al (2003) do not only provide an overview of many different personalization techniques but also of how they were selected for the project in hand.
Conclusion Current state of development A number of different approaches to personalization are currently used in various applications. Highly sophisticated e-commerce sites usually employ a combination of multiple methods and contain adaptive as well as adaptable elements in a coherent userinterface. As the case study of Amazon.com has shown, organisations can add real value for the users by making it more convenient to tailor the presentation to individual needs. Adaptive methods for personalization are very powerful means to manage information overload and simplify user-interfaces. However, they are not frequently implemented in
10
large-scale systems. Despite of highly developed algorithms, they can produce unpredictable and unwanted results. This is an area where further research is necessary, as adaptable methods have the serious disadvantage that their configuration remains static until the user changes it explicitly. This can be a tedious task for most of the users which lowers the value even of the most flexible customization services. Hence there are only a small number of users who actually want to customize their pages. Directions for future research & development Improve group based personalization The personalization based on a user’s preferences is increasingly extended by collaborative methods. Instead of recommending items based on the profile of a single user, the system should try to take advantage of other user’s ratings and preferences as well. While this might in fact increase the accuracy of predictions it raises the issue of proper user selection. New Input methods A lot of research is done to improve the communication between users and computer systems. Among them are methods that allow formulating queries in a natural language rather than in special “inhuman” query syntax. As the system has a model of the domain and the necessary vocabulary it can ask “intelligent” questions if the user made unclear statements (Zadrozny et al., 2000). No matter whether natural language commands and queries are spoken or typed, they can make the use of personalized websites, speech recognition based telephony services and other information systems much more individualistic. Combination of explicit and implicit data Both adaptive and adaptable methods have their strength and weaknesses and may be more or less applicable for a particular situation. However, it seems to turn out, that the combination of explicit and implicit user data provides the best results. On one hand the effort of manual customization is minimized and on the other hand an adaptive method will not cause much unpredicted results when it is limited by explicit statements. Wells and Wolfers (2000) explain how customers who use online banking services need to make some basic statements about their financial goals and situation. After that, adaptive methods are used to offer financial products and services which are tailored to the customer and suitable for his particular needs. However, not all techniques applicable to a certain scenario will be successful in practice. Alpert et al (2003) show in their study about user’s attitudes towards adaptive systems that users have a strong desire to always be in full control of all interaction. It is therefore important to carefully analyze the potential acceptance barriers of a designed solution before finally deploying it. Future Outlook Personalization technologies have found their way out of experimental systems of researchers into commercial applications. They are powerful means to handle information overload, to make complex information systems more usable for a heterogeneous group of people and help online businesses to establish personal relations to their customers
11
(one-to-one marketing). Although we have focussed on applications of personalization to websites, they can be used in wide range of human-computer interactions. Personalization techniques are the key to “mass-customization” and provide people with a much more individual experience instead of standardized services. Within the next few years, even more systems will incorporate these technologies and things like personalized speechrecognition services, which seem to be futuristic today, will soon be reality.
Bibliography Alpert, S. R., Karat, J., Karat, C., Brodie, C., & Vergo, J. G. (2003). User attitudes regarding a user-adaptive ecommerce Web site. User Modeling and User-Adapted Interaction, 13(4). Belkin, N. J. (2000). Helping people find what they don’t know. Communications of the ACM, 43(8). Hirsh, H., Basu, C., & Davison, B. D. (2000). Learning to personalize. Communications of the ACM, 43(8). Kantor, P. B., Boros, E., Melamed, B., Meñkov, V., Shapira, B., & Neu, D. J. (2000). Capturing human intelligence in the Net. Communications of the ACM, 43(8). Karat, C. M., Brodie, C., Karat, J., Vergo, J. G., & Alpert, S. R. (2003). Personalizing the user experience on ibm.com. IBM Systems Journal, 42(4). Manber, U., Patel, A., & Robinson, J. (2000). Experience with personalization on Yahoo! Communications of the ACM, 43(8). McCarthy, J. (2000). Phenomenal data mining. Communications of the ACM, 43(8). Mobasher, B., Cooley, R., & Srivastava, J. (2000). Automatic personalization based on Web usage mining. Communications of the ACM, 43(8). Mulvenna, M. D., Anand, S. S., & Buchner, A. G. (2000). Personalization on the Net using Web mining. Communications of the ACM, 43(8). O’Connor, M., Cosley, D., Konstan, J. A., & Riedl, J. (2001). PolyLens: A recommender system for groups of users. Proceedings of ECSCW 2001, Bonn, Germany. Pine, B. J., II. (1993). Mass customization. Boston: Harvard Business School Press. Riecken, D. (2000). Personalized views of personalization. Communications of the ACM, 43(8). Schafer, J. B., Konstan, J. A., & Riedl, J. (2001). E-commerce recommendation applications. Journal of Data Mining and Knowledge Discovery, 5(1/2). Wells, N., & Wolfers, J. (2000). Finance with a personalized touch. Communications of the ACM, 43(8). Zadrozny, W., Budzikowska, M., Chai, J., Kambhatla, N., Levesque, S., & Nicolov, N. (2000). Natural language dialogue for personalized interaction. Communications of the ACM, 43(8).
12
Terms and Their Definition Adaptable Personalization Systems are systems which can be customized by the user in an explicit manner, i.e. the user can change content, layout, appearance etc to his needs. Adaptive Personalization Systems change the presentation implicitly by using secondary data. This data can be obtained from a variety of sources, for example from the users actions, from the behaviour of other users on that site or based on the currently displayed content. Decision Support Systems/Tools in a wider sense can be defined as systems/tools that affect the way people make decisions. Mass Customization refers to the customization of products and services for individual customers but at a mass production price. Personalization is an approach to increase the usability of complex information systems and present the user a comprehensible interface which is tailored to his/her needs and interests. Recommender systems are a special type of decision support systems which give recommendations for further actions or related items. Tailoring in the context of personalization can be with respect to content to navigation. Content refers to the information being displayed and navigation refers to the structure of links which allow the user to move from one page to another.
13
