Development of Website Design Personalization Service. Using Design Recommender System. Jong-Hwan Seo*, Kun-Pyo Lee**. *Tongmyung University of ...
Development of Website Design Personalization Service Using Design Recommender System Jong-Hwan Seo*, Kun-Pyo Lee**
*Tongmyung University of Information Technology, Dept of Computer Graphics, 535 Yongdang-dong, Busan, KOREA, [email protected]
** Korea Advanced Institute of Science and Technology, Dept of Industrial Design, 373-1 Guseong-dong, Yuseong-gu, Daejeon, KOREA, [email protected]
Abstract: Personalized services in websites continue to increase with a distinct tendency toward the user-oriented paradigm; however, current personalized services related to website design are based on a very simple type of personalization, such as merely showing assembled webpages to a user according to his/her selections out of various design alternatives. Such a type of personalization has two problems; 1) the process of making selections imposes a heavy burden on users; 2) it is difficult to respond effectively to the user's preference, which is quite complicated and changed frequently. This study aims to improve the current personalized service in website design by making it more user-friendly and more user-adaptive. We began this task by first reviewing the present situation of personalization service in website and then attempting to apply a recommender system to the field of website design for more user-adaptive personalized service. Through literature review, major attributes of recommender systems from several points of view were identified such as recommender contents, user information, and recommender method. This literature review was followed by an analysis of how to apply these features of recommender system to website design efficiently. Based on these investigations, a design recommender system was proposed based on a collaborative filtering method that produces personal recommendations by computing the similarity between a user's preference and the one of other users. Our design recommender system consists of four major modules: User Entry Module, Design Attribute Entry Module, Design Customization Module, and Design Recommender Module. Finally, the detail procedure and methodology for the practical use of this system was developed. Key words: Web Design, Recommender System, Personalization
1. Introduction The need for personalized service on website has been increased these days, but personalized service related to website design has stagnated at the level of merely assembling design elements in accordance with the user's selection. Future website design personalization services must improve to lighten user's burden in the process of personalization and produce results that are more adaptable. This study examines various personalization services, which are used successfully in other fields of website development and then suggests a design recommender system using collaborative filtering techniques. The recommender system suggested in this study consists of four modules, such as user entry, design elements entry, design customization, and design recommendation. Finally, we made a more detailed procedure and structure for the practical use of our system.
2. Personalization Service in Website Design 2.1 Overview Current personalization services in website design are based on user’s voluntary participation and active selection. If users want to experience personalized page design, they should express their preference about page design elements such as layout, color, image, and typeface. Then, the personalization system simply assembles design elements selected by the user into a webpage and displays the result. Exite.com is a typical example of this type of personalization service (Figure 1).
Figure 1. My Exite (http://www.exite.com)
2.2 Problems This type of personalization design service has some limitations: 1) it requires many procedures and detailed tasks, imposing a heavy burden on the user; 2) it primarily depends on the user’s voluntary offer of information, so it cannot work well for the users who do not exactly know what they want or cannot explain it thoroughly. To solve these problems, the personalization design system needs to be improved to lessen user’s burdens and understand the user’s behavior and taste by using user-adaptive techniques . Actually, many advanced websites have already made practical applications of user-adaptive techniques to their personalization service. For instance, the world’s largest online shopping site, Amazon.com, has been introduced an exemplary site which realizes user-adaptive personalization service. Amazon.com recommends various contents to the user by analyzing his/her shopping history and clickstream data. To realize this user-adaptive personalization service, it needs to study techniques to recommend content that a user is likely to prefer. These techniques have been studied vigorously as a “recommender system.” 3. Personalization Recommender System 3.1 Overview and Constitution of a Recommender System A recommender system is basically a system that can learn about a user’s personal preference based on the user’s characteristics and behaviors and can then provides the most appropriate content to meet the user’s needs. Recommender systems have been applied to various websites and recognized to be useful in recent years. For instance, Netflix.com, a DVD rental site in the United States, developed their recommender system, called Cinematch, which is a service that suggests DVD titles a user may like to rent based on the user’s own tastes and the tastes of past users. Neflix.com has become the world's largest DVD rental site due to its personalization service using Cinematch.
A recommender system consists of three elements as shown in Figure 2. First, various recommendation content which is presented to users has to be made. Then, users’ preferences or behavioral data on these contents must be gathered. Finally, it needs to choice type of recommendation technique about how to analysis these user data and select the optimal content to each user .
Figure 2. Constitution of Recommender System
3.2 Types and Features of Recommender Systems Most Recommender systems adopt two types of recommendation techniques: a content-based approach and a collaborative filtering approach. Content-based filtering is based on content analysis of the considered objects and its relation to the user’s preferences. For content-based filtering it is therefore necessary that the results of content analysis and user preferences can reliably be determined. Landsend.com, the leading clothing company of the US, has adopted a content-based approach to their website service (Figure 3). In their recommender system (“My Personal Shopper”), users are categorized into various types based on their preference for a series of clothes displayed in website. Then, the recommender system analyzes Landsend’s products and recommend what each user would like. This type of recommendation is based on analyzing and classifying contents; therefore, it is difficult to apply this approach to multimedia content due to the limitations of current content analysis technology .
Figure 3. Content-based approach: My Personal Shopper (Lansend.com)
Collaborative filtering is another approach to identify content that is relevant to a user. The collaborative filtering approach identifies other users that have shown similar tendency to the given users and make a recommendation based on those similarities. Generally, in the collaborative filtering, the content analysis is ignored and only other user’s opinions on the considered content are considered relevant. Therefore, the collaborative filtering approach
is especially interesting for content for which content analysis is weak or impossible. However, the performance of the collaborative filtering approach relies on the available user preference data for the considered content and therefore fails when few or no opinions are known. Table 1. Types and Features of Recommender Systems
Collaborative filtering approach
Recommends based on content analysis and the given user’s past preference data
Finds other users that have shown similar tendencies to the given users and recommends what the past users have liked
Content can be recommended without any preference data from the users.
Contents analysis is ignored and only the user’s preference data on the considered contents are relevant
It is difficult to apply to content that is hard to analyze and classify.
It is difficult to apply to content which have few or no preference data from the users
My Personal Shopper (http://www.landsend.com)
Amazon’s New Recommendation (http://www.amazon.com)
4. Design Recommender System for Website Personalization Service 4.1 Basic Framework The main features and requisites that are needed to build a recommender system for website design are examined from the point of view of its three components.
· Recommendation Content A website consists of many webpages; therefore, webpage design can be used as recommendation contents for the design personalization service. Webpage design has various elements such as text, graphics, color, layout, etc. In current personalization design service (see Section 2.1), users select their favorite elements out of many alternatives made by designers, and then the personalization system assembles these preferences into a page design. This cooperative approach is also very practical in making content recommendations because it is impossible for a few designers to make the whole page design needed for recommendation. Therefore, our recommender system utilizes this cooperative approach of current personalization design service to make page designs, and the results are used as recommendation content. · User Data User data are divided into two classes: explicit user data and implicit user data. Explicit user data are gathered by asking users to give their own preferences by filling out a form that requests particular information. On the other hand, implicit user data are obtained by analyzing user’s usages and behaviors in the given website . For instance, Amazon.com tracks a user’s purchases and makes recommendations based on his/her purchase history. Generally, implicit user data are more sensitive and elaborate than explicit user data; however, techniques related to gathering implicit data are in their initial stages and require a great deal of money and facilities, such as tracking devices. Therefore, our recommender system is based on collecting explicit user data by means of the user’s direct participations, such as the user’s rating. · Recommendation Technique Website design consists of various multimedia elements that in themselves are interactive and have complicated
features. Since website design has a subjective character, its evaluations vary with the personal tastes of users. Accordingly, it is very difficult to analyze and classify website design objectively. Even if it could be done, the result would be unacceptable in most cases. Therefore, it is very hard to analyze and classify page designs into categories by using a very formative framework such as a content-based approach and recommend page designs according to the result. On the other hand, if we apply a collaborative filtering approach to design a recommender system, we would expect to find a user group that has the most similar preference to the given user by analyzing user’s preference data on specific page designs without classifying them and then recommend the most preferable design to each user. For this reason, the collaborative filtering approach is an appropriate technique for recommendation of subjective contents like page design; however, the collaborative filtering approach has weak points that must be improved. For instance, it is difficult to recommend appropriate contents to a user who is newly registered because the user has no preference data at that time (known as the “new user problem” or “cold-start problem”) . To solve this problem, our recommender system applies another type of user data such as demographic data in the case of a newly registered user. Figure 4 summarizes the three main issues that the recommender system suggested in this study focuses on.
Figure 4. Main issues of the Design Recommender System
Figure 5. Structure of the Recommender System for Website Design Personalization
4.2 Structure and Procedure The recommender system for the website design personalization suggested in this study consists of four main modules based on a personalization database (Figure 5). The contents and procedures of each module are described as follows.
1) User Registration Module Those who want to use the recommender system for website design personalization can register themselves as new users by furnishing general information about themselves. The data requested in this module mainly consist demographic and background information on each user. All the collected user data are stored in a user profile database of the recommender system to be used in following modules. 2) Design Elements Registration Module The design elements registration module aims to prepare the groundwork for developing a wide range of content for personalization. By using this module, designers can divide page design into appropriate constituent elements and register their various design alternatives for each element in a design element database. Registered design alternatives are used as basic ingredients with which users can make their own page designs. 3) Design Customization Module The design customization module basically has a similar procedure to the current personalization service in website design reviewed in Section 2. However, it is improved in two aspects as follows: first, a page design assembled by a user is registered in a design content database, which can be reviewed and used by another user as content for personalization; second, whenever a page design is registered, reviewed, or used by a user, the user rates it. Users can express preferences by rating page design presented to them. Therefore, the recommender system can have users’ preference data on each page design, and these preference data are used as a basis for recommendation in the following module.
Figure 6. Procedure of the Design Customization Module
4) Design Recommender Module In the design recommender module, we adopt the collaborative filtering approach to analyze similarity among users and identify the nearest neighbors of each user by using their preference data. For this purpose, all of the
users who makes their own page designs in the design customization module or receive a page design recommended by the system should offer their preferences to the page design by filling out a rating form, and the results are stored in the user profile database. The design recommender system finds users who have the most similar preference pattern to the given user in the user profile database and then searches and retrieves page designs that they give a high rate of preference to in the design contents database. In the case of a newly registered user, this recommender system uses demographic data instead of preference data in order to solve the new user problem as stated above. Therefore, for a new user, a page design that demographically similar users have preferred will be recommended. By using the procedure outlined above, a page design that meets the individual preference and taste of each user can be predicted and recommended.
Figure 7. Procedure of the Design Recommender Module
5. Discussion Our personalization service proposed in this research is designed to improve the problems of current personalization design service, which are investigated in Chapter 2.2, and we suggest a more efficient approach as follows: x User data can be accumulated and utilized through the recommender system, which has not yet been studied and
adopted for personalization design service, and therefore the heavy load of works imposed on users in the current service can be reduced significantly. x The current personalization design services provide page designs only according to the user’s explicit demands that include detailed specification of components. However, our newly developed service is based on adaptive user modeling technique with user’s previous response and can predict their preference, which even they do not know, and suggest page designs that they are likely to prefer. Moreover, in terms of applying recommendation technique, personalization service of this research is quite different from the previous recommender systems that we find in the current practice. x Existing recommender systems have been developed for the fields that are easy to analysis because their contents are based on text data. However, the recommender system proposed in this research has focused on website design that has more complicated contents like images. x There are many kinds of recommendation techniques. Most existing recommender systems have been developed by focusing on one of these techniques. Our system also uses a collaborative filtering as a main recommendation technique, but include a demographic filtering in the recommendation process for improving performance. It is a hybrid recommender system with a staged process in which demographic filtering technique is employed first. If user preference data are collected with sufficient confidence, then a collaborative filtering with preference data is attempted. Through this switching hybrid procedure, our system provides the ability to solve the ‘new user’ problem that most collaborative filtering recommender systems suffer from.
6. Conclusions and Future Research The need for personalized website design has increased in recent years. The current approach for personalized website design has been easily applied to websites due to their cost-effective features, but the current approach cannot easily provide a more refined personalized service because of its lack of a user information database. In this study, the design recommender system is investigated as a more advanced method for website design personalization. We outlined current recommender systems and identified recommendation techniques, especially collaborative filtering, as a key technology for the creation of user-adaptive personalization service in websites, which produce personal recommendations by computing the similarity between a user's preference and the those of other users. Then, we examined problematic issues for the collaborative filtering technique, such as the case when a new user with no or very few ratings cannot be reliably matched against other users. To solve this new user problem, we use demographic data instead of preference data; therefore, a page design that demographically similar users have preferred will be recommended for a new user. Based on these investigations, we presented a framework of a design recommender system for website personalization based on the collaborative filtering approach. Our design recommender system consists of four major modules: User Entry Module, Design Attribute Entry Module, Design Customization Module, and Design Recommender Module. Finally, we closed by proposing the detail procedure and methodology for the use of these modules. Our work has limitations and directions for further research as follows: x This study aims to suggest a basic concept and prototype to improve the current personalization service for website design by using exploratory research methods such as a literature review and a case study. Therefore,
in-depth technical studies should follow for the more practical use of the proposed prototype. x Our recommender system is based on collecting explicit user preference data by means of user’s direct participations such as user ratings; however, implicit user preference data obtained by analyzing a user’s usage and behavior should be mixed with explicit user data to build a more sensitive and effective recommender system and support lifecycle personalization. Therefore, our future work in this area is to include the implicit user preference data in our recommender system. User-adaptive services and recommender systems in website have developed in response to a manifest need: helping users deal with the world of content and services abundance and overload. We believe that the collaborative filtering approach will be a powerful tool for the construction of user-adaptive personalization service in website design.
References 1. A. Kohrs, B. Merialdo, Creating User- Adapted Websites by the Use of Collaborative Filtering, Interacting with Computers, 13, 697-698 (2001). 2. L. Terveen, W. Hill, Beyond Recommender Systems: Helping People Help Each Other, in HCI In The New Millennium, Jack Carroll, ed., Addison-Wesley, 3-4 (2001). 3. C.-H. Lee, Y.-H.Kim, P.-K. Rhee, Web personal expert with combining collaborative filtering and association rule mining technique, Expert System with Applications, 21, 131 (2001). 4. Keith Instone, Information Architecture and Personalization, Argus Associates (http://argus-acia.com/white_ papers/personalization.html), 6-7 (2000) th
5. Philip S.Y., Data Mining and Personalization Technologies. In Proceedings of the 6 International Conference
on Database System for Advanced Application, Taiwan (1999) 6. Robin Burke, Hybrid Recommender Systems: Survey and Experiments, User Modeling and User-Adapted Interaction, 12, 331-370 (2002)