Wallin, Erik, Lennstrand, Bo & Persson, Christian: "Personalization Matured - Rediscovering the Essence of One-toOne Marketing". The IADIS e-Commerce 2006 conference, Barcelona, Spain, 9-11 December 2006

PERSONALIZATION MATURED - REDISCOVERING THE ESSENCE OF ONE-TO-ONE MARKETING

ABSTRACT The paper is about personalization system applications in different kind of settings. The studied systems utilize descriptions of information characteristics to perform content based filtering in order to increase stated business objectives. Data and experiences from action research, covering three different product areas - sports retailing, online newspaper and online recruitment sites, are presented and discussed, as well as the considerations made when developing and adapting a personalization system for each of these areas. Data describing personalization efficiency have been collected through an own-developed method, bifurcation, which makes it possible to compare the behavior of personalized groups with the behavior of an unpersonalized control groups. The results show that personalization has a positive influence on the stated business objective – traffic increase. KEYWORDS Personalization, One-to-one marketing, E-commerce, Bifurcation, Relationship marketing, Non-intrusive

1. INTRODUCTION Since 1998, the Kimono-project has worked with the development of personalization systems for web sites in different settings, and with an own-developed tool - bifurcation - to measure the effectiveness of such systems. This paper presents some experiences from the project; also discussing problems personalization efforts may meet in varying environments. The paper deals with successful implementations of personalization systems as well as with those who failed. The research area is consumer organization interactions that utilize a technical system. This can for instance be an Internet web-site dialogue, a creditcard payment or a phone call. Though the main objective is personalization of web site content, the paper addresses a broader area of marketing, where the problem of individual information adaptation also is relevant.

2. OBJECTIVES This study concerns problems and solutions related to the creation, implementation and performance of personalization systems on three different kinds of web sites: • The Sport Planet. A personalized web site developed for one of Sweden’s largest independent sports retailers. • Svenska Dagbladet. Personalization system for the website of Sweden’s third largest newspaper. • Offentliga Jobb. Personalization system for a recruitment site. Some experiences from system implementations in these three business cases will be presented and discussed.

3. METHODOLOGY The research approach can be characterized as action research. Initially, the project started with the ambition to collect data about personalization methods and personalization efficiency from companies using personalization technique. To obtain a fully controllable research tool, a personalization system, integrated with a module for measuring the personalization efficiency was developed (the bifurcation module). (Persson et al 1999)

The bifurcation module divides visitors to a website stochastically into two or more groups with different personalization configurations (figure 1). Thereby the system is able to present different content to randomized groups of visitors whose responses are recorded and measured separately (double-blind test). By giving one of the groups a configuration that deactivates the personalization algorithms, a control group is created. An identification system and a visitor database ensures that a certain visitor will be directed to the same group every time (s)he visits the website. The bifurcation system can also be used in a trial-and-errorlike way to try different personalization configurations in order to find the best parameter settings when optimizing the effects of personalization.

Figure 1. Schematic view of the bifurcation function.

4. CONCEPTUAL FRAMEWORK 4.1 Relationship marketing In all kind of relationship-based marketing - whether it is named Direct Marketing, Relationship Marketing, CRM or One-to-One Marketing - the company should act on information known of the customer (visitor), e.g. about products bought, preferences, etc. This is a way to deal with customer attention as a scarce resource. A consumer can only take into account a minor part of the commercial and public messages (s)he is exposed to. Probably the consumer may notice messages that fit his/her interests and neglect those who do not. Relationship-marketing methods rely to some extent on information from computerized business system. Some company-customer interactions and marketing measures are fully automatic, while others involve people to a great degree. Two examples to illustrate this: Volvo card – a fully automatic process: If a Volvo card owner has not used the card for three months, (s)he will get a discount offer on the next monthly bill. The card owner is identified by a unique identifier (the card) in a way most people perceive as non-intrusive. The card owner is differentiated from others automatically by a rules-based system, which monitors the last month's bills. Card uses are automatically registered and remembered by the billing system and the interaction (bill via mail) is automatic and batchlike, customized automatically and direct by an added offering text on the bill. Phone support – mainly manual process: In many phone support systems, the consumer has to enter some kind of unique identifier (e.g. order number, social security number, bank account, serial number, etc.) in an intrusive way. Stored data about the consumer's previous interactions, which differentiates him/her from others, are automatically shown on the adviser's screen. From this the adviser can manually customize his behavior in the ongoing interaction, and manually make those notifications that should be remembered in future.

4.2 Personalization models Automatic individual information adaptation (personalization) has for long been proposed as a solution to the problem with information overload. Peppers et al. has suggested a model for one-to-one marketing including four steps, which proposes a way to structure communication with individuals: Identify, Differentiate,

Interact, Customize (Peppers et al. 1999). Feurst proposed a slightly modified version of this model, putting Differentiate together with Identify and adding Remember (Feurst 1999). Foster et. al. proposed a cyclic model (figure 2), consisting of five phases: Store, Cache, Mine, Tune and Target (Foster. et al. 2000). This last model captures the continuous process of adapting content to a customer’s interests, both from what is registered about his/her behavior in previous contacts, and what is shown in the ongoing session.

Figure 2. The personalization cycle (Foster et al., 2000).

In figure 3, we suggest a model, which is a combination of the mentioned models. Identification is needed in the beginning of every visiting session, but is not a part of the following cyclic process. When identified, visitor data is retrieved from the memory stored in a visitor database (e.g. visitor profile). In the beginning of a session, there is only previous (or no) behavior available to govern the differentiation method. The content to be personalized is rated for each visitor (or the visitors are rated, or fits into some segment model). In the customize phase, algorithms govern the selection of relevant content to present. The personalized content is presented in the interaction phase and the visitor leaves traces pointing at his current interests. After the interaction phase, the memory is updated to include the visitor's new actions, and the process continues with a new differentiation check, and so on and on until the session ends. An issue to deal with in this process is what weight should be taken to the visitor's historical behavior versus his/her behavior in the ongoing session.

4.3 Modus operandi Personalization could be described as the process of filtering a set of items (products, information) through an interest profile of a customer. Therefore a personalization system should contain both a method for creating customer profiles and a method for filtering1.

1 In the simplest form of personalization a cookie is used to store information from the visitors previous sessions. This can simplify the visitors navigation on the site, when (s)he wants the same content as last time. This requires neither a personal profile nor a method for filtering, but can still be efficient in some situations, Using the 1-1-model, the cookie acts as identifier, differentiator and memory to help customizing the site.

Differentiate Content rating

New visitor

Retrieve previous behaviour

Identify

Memory

Customize

Store ongoing behaviour

Content selection

Interact Customer input

Figure 3. Model of the personalization process.

In an Internet environment, visitor profiles can be created either by asking the visitors to define their preferences (explicit) or by monitoring the visitors' behavior on the site (implicit). The former approach requires active participation from the visitor, while the other approach is entirely non-intrusive. Identification could rely on a unique identifier (e.g. bank account number, logon and password, email address, etc.), or on some kind of approximate identifier (cookie, phone-number, ip-number, etc.). The identification process can be intrusive (e.g. when asking a website visitor to log on) or non-intrusive (e.g. by using a cookie). In many situations, intrusive identification will make the visitor leave the site. Remember means saving interaction data for future use, in the ongoing session as well as in forth-coming ones. This can be done automatically or by some kind of manual registration by the visitor or by an operator. Differentiation could be automatic or to some extent founded on manual coding (e.g. meta-tagging). The differentiation could be based on the consumers’ explicit preferences or his/her earlier behavior (implicit). Popular approaches for implicit differentiating are either deterministic and deductive such as rules-based algorithms or probabilistic and inductive such as content-based filtering and collaborative filtering (Rijsbergen 1979). The content-based approach provides items that are similar to what the visitor has favored in the past. The content-based approach requires that the items be pre-categorized according to schemata. In the collaborative filtering approach, other visitors that have showed similar preference to the given visitor are identified and the system then provides what they like. Customize means presenting offerings or information objects suited to each visitor's preferences as they have been expressed explicitly by registered preferences and/or implicitly from the ongoing and previous interactions. This process can be automatic or manual. Interaction could be automatic or personal. Interactions can take place in real-time or through batch-like processes (e.g. questions and answers via e-mail). When summing up the conceptual framework, a division or polarization of concepts significant to personalization could be: Automatic

Manual

Explicit

Implicit

Intrusive Deterministic

Non-intrusive Probabilistic

To what extent is manual operation demanded by the system’s operator. Whether the personalization is explicitly chosen by the end-user or if the personalization is automatic from an end-user’s perspective. How the end-user experiences the personalization. The inference model for personalization.

The personalization system used in the case studies is a Java-based application that serves the companies multiple front-end web servers. For the behavioral collection the web servers report visitor id and article number to the personalization system asynchronously. The personalization process is carried out synchronously by the web server sending the visitor id and a set of articles to choose from, the result (the personalization) is returned to the web server. In the case studies the companies’ means of user identification (cookies in both cases) were used. The personalization system used in the case studies can be summarized as: automatic, implicit, non-intrusive, inductive, probabilistic, and content based.

5. BUSINESS CASES Three business cases with personalizing functionality developed in the project have been studied. The cases differ with respect to how much manual handling the personalization system demands. The Sport Planet Case demands a high level of manual handling of product presentation and categorizing. In the Svenska Dagbladet case the categorizing is funded on a standardized metadata descriptions but all content has to be manually described ac-cording to this standard. In the Offentliga Job case the personalization process is fully automated where the service already is categorized according to a geographical classification and a job categorization standard.

5.1 The Sport Planet The case object was an independent sporting goods retailer in Stockholm, Sweden, with no activities on the Internet before the time of this study. In 1999, the company wanted to start sales on the Internet. An ecommerce system with personalization functions was developed. However, the e-shop never opened due to circumstances discussed below. The e-commerce system was constructed to support real-time identification, selection, and content creation with the added functionality of bifurcation. It was decided not to use operational methods when creating consumer profiles. Instead, the profiles are generated from the products, which a visitor looks at or buys, through a rule-based system. However, the buying behavior associated with sports products makes it difficult to estimate a visitor’s interests in new products from what he has previously bought: For instance, does the purchase of three bicycles during the last year mean that you have all the bicycles you need - or that you are especially focused on bicycles, interested in all information about bicycles, and will continue to buy them? As a solution to this problem of identifying long-lasting customer attitudes and preferences, a differentiation function that uses value-carrying parameters associated with the articles was developed. Each articles was manually classified ac-cording to five criteria: the judged 1/ Strength of brand, 2/ Price and quality level, 3/ Fashion and actuality level, and whether the article was suited for a certain 4/ sex, or 5/ age. When a visitor buys or looks at an article, his profile is updated by the article’s specific parameters. E.g. if you look at a high priced product, your profile is updated, showing some interest in high-price articles. In this way, the customer's selection of products with their inherent values is interpreted as a reflection of the customer’s preferred personal values. To handle the important question of how to weight the influence from ”what you are looking at” in relation to ”what you buy”, the system is adaptable, providing the opportunity to change weight factors through an administrative interface. The case object was a rather small company (about 40 employees) with limited re-sources. Their computer system was old and incomplete. Article numbers and descriptions could not be transmitted directly to an e-commerce system. Every product had to be de-scribed manually - a considerable job since the company is marketing about 10 000 articles, especially since there was no pictures to show on the website unless products were photographed. Furthermore, e-shop ordering, billing and payment routines in the e-shop could not automatically be transferred to the existing system. Although the web shop technically was ready to launch in autumn 1999, the company could not decide whether to do so or not. The company doubted that the estimated sales on the Internet could compensate for the costs. So the Boo.com-crash happened, and the ideas about Internet sales was abandoned. The main conclusion from the Sport Planet Case is that personalization of web sites can be hard to obtain if the process demands much extra manual handling regarding product presentation and categorizing.

5.2 Svenska Dagbladet (SvD) A version of the personalization system, adapted to news services, was tested on the web-site of Svenska Dagbladet, Swedens third largest newspaper, in 2001. Examples of personalization of news services on the Internet reach back to early experiments in the early 1990s [6]. A popular method of adapting information is content filtering, which refers to the concept of selecting content for a customer based on semantic descriptions of the news items and corresponding customer profiles. Content filtering was used in the implementation at SvD. The personalization operation is based on two metadata descriptions of the news material, 1/ Category (e.g., sports, entertainment, international news) and 2/ Geographic location (e.g., Stockholm, Sweden, Europe, etc). For the category description a TT NITF metadata structure was used (Stadler et al. 2003), originating from the International Press

Telecommunications Council’s News Industry Text Format. A journalist who has written an article also encodes it with metadata. SvD’s business objective behind the personalization system implementation was to in-crease the average number of page impressions (defined as the total number of downloads of web pages) a customer demands each session. One of the requests was that the first page should contain one personalized area in order to address the problem that the majority of the customers left the site after only viewing the first page. A second personalized area follows at the end of an individual editorial article. During one week from November 19 to 25, 2001, a test was performed. Altogether 81,305 test subjects were registered and divided into four heuristically chosen bifurcation test groups (three personalized and one non-personalized, each with 25 % of the traffic). The three personalized groups differed regarding the weight given to the visitor's historical behavior versus his behavior in the ongoing session. In one group, historical behavior was favored (33% weight for current session behavior, 67% weight for history). In the second group, session behavior was favored (67% versus 33 for history). In the third group, historical behavior was neglected and personalization was only governed by what happened in the ongoing session. The customers in the control group received articles selected alphabetically rather than based on individual personalization. During the test all interactions performed by the customer were logged. The results show that the 33 % session weight personalization group was on average 1 % better than the non-personalized reference group with regards to average number of interactions per session. The 67 % session weight group was 10 % better, and the 100 % session weight group was 8 % better than the reference group. All differences between the reference group and the test groups were found to be statistically significant when carrying out the ANOVA test. Kruskal-Wallis’ test also showed a significant group difference. A problem in the SvD case was that the personalization system did demand a more precise categorization of news articles than needed for other use within the company. Because of this, all articles were not categorized with full relevance.

5.3 Offentliga Jobb This study concerns one of Sweden’s largest Internet-based recruitment services, Offentliga Jobb (www.offentligajobb.se). The mission of the Offentliga Jobb online recruitment ser-vice is to present job offerings to people searching a new employment within the public sector. The recruitment service company had their website personalized in 2002. Offentliga Jobb’s business objectives with the personalization system are to increase the traffic on the website and increase ease-of-use and visitor satisfaction. In the

Figure 4.Offentliga Jobb website. Personalized sections on the first page

study data from the use of the personalization system during one year (2004) is analyzed. (The personalization system is still running at Offentliga Jobb.) The base for the study was an empirical material, collected during a two-year period, covering more than 3 million unique visits. Time- and content-dependent variations of the visitors’ behavior patterns when exposed to information adaptation have been analyzed. The personalization system implementation at the Offentliga Jobb uses two different visitor identification mechanisms: anonymous visitors get a cookie set upon his or her computer as a unique identifier, while

registered visitors are identified through the log-in procedure. However, only an almost insignificant part of the visitors chose the login procedure. The visitor's profile is built from the visitor's choice of job offerings. All interactions are collected and saved as the visitor’s individual profile. A content-based filtering method is used, i.e. the information adaptation process is based on descriptions of the content. Of-fentliga Jobb’s content-based filtering method uses only two dimensions to characterize the job offerings, on one-hand metadata profiles for describing the job-offerings in a standard format called SSYK, on the other geographical classifications. All job-offerings have to be classified according to this system. This means that the categorization needed for personal-ization is a fundamental part of the routines at OJ, and not anything extra needed to make the personalization system work. The following personalization functions were developed in cooperation with Offentliga Jobb: 1. First web page navigation (ellipse A in figure 4) 2. First web page selection of job offerings (ellipse B in figure 4) The results from the Offentliga Jobb case show that this simplest possible form of personalization seems to be enough to increase the tendency to continue with more clicks. Among visitors interacting with one more page than the opening page the personalized group is 9,9 % larger than the control group. To obtain the same amount of clicks without the personalizing function installed the number of visitors at Offentliga Jobb should need to be in-creased with 7,2 % or 133 507 visitors. The value of personalization in this setting therefore could be compared to the costs and efforts needed to attract another 130 thousand visitors.

6. DISCUSSION Empirical data that support the efficiency of the personalization technology, i.e. on ROI, functionality or usability, are rarely found in literature. Most research on this issue seems to be based on investigating small groups of respondents in experimental settings (Chesnais et al. 1995). In our research, a personalization system with built-in methods for measuring visitor responses has been used to verify the impact of personalization. In our studies we have measured the im-pact of information adaptation in a newspaper environment on the Internet and one on a recruitment service on the Internet - both showing a statistically significant increase in web-site traffic when using the personalization system (Yu Li et. al 2005, Kohrs et al 2001, Bradley et al 2003). In both cases the on average increase in traffic when personalizing the content is about 10 %. The evaluation function of the information adaptation system used in our research takes the business operator’s standpoint when pragmatically focusing on a few measurable busi-ness objectives. This means to measure the very operation of a personalization system. To carry this out, the method we name bifurcation is essential. A parameter with importance for the discussion of the efficiency of personalizing systems is how much treatment the content has to undergo to be ready for a personalizing process. This will of course vary between different business categories and companies. However it is obvious that a too complex manual handling process as was the situation in the Sport Planet case will prevent the implementation of personalization functionality. The easiest situation is where the content already are categorized and the necessary parameters for a personalizing process are few, as in the Offentliga Jobb case study. One interesting finding in the SvD case was that the visitors’ behavior in ongoing sessions should be given more weight than their historical behavior in order to maximize the operator’s objective.

7. CONCLUSION The case studies presented in this paper utilize descriptions of information characteristics to perform content based filtering in order to increase stated business objectives. Data and experiences from action research, covering three different product areas - sports retailing, online newspaper and online recruitment sites, has been presented and discussed. Data de-scribing personalization efficiency have been collected through an own-developed method, bifurcation, which makes it possible to compare the behavior of personalized groups with the behavior of unpersonalized control groups. The results show that personalization has a positive influence on the stated business objective.

REFERENCES Bradley & Smyth, 2003. Personalized information ordering: a case study in online recruitment. Knowledge-Based Systems 16 pp. 269-275 Chesnais et al. 1995. The Fishwrap Personalized News System. IEEE 2nd International Workshop on Community Networking Integrating Multimedia Services to the Home, Princetown Feurst, O., 1999. One-to-one marketing – filosofi och metod.. Liber Ekonomi, Stockholm. Foster, C.et al, 2000. The Personalization Chain. Demystifying Targeted Delivery. Jupiter Communications. Site Operations/Vol. 3. Kohrs & Merialdo, 2001. Creating user-adapted Websites by the use of collaborative filtering. Interaction with Computers 13 pp. 695-716. Peppers, M. et al, 1999.The one to one fieldbook: The Complete Toolkit for Implementing a 1to1 Marketing Program. Doubleday, New York. Persson, C., et al, 1999. COTIM99 Proceedings. Electronic Commerce: Behaviors of Suppliers, producers, Intermediaries & Consumers, Volume 3. University of Rhode Island. The SportWeb Case: Design and Prototyping of a System for E-Commerce and Research. Rhode Island. USA. Rijsbergen, C. J. van. 1979. Information Retrieval. 2 Rev Ed edition. Butterworth-Heinemann, London. Stadler, H. & Lindgren, J. Dec. 8 2003 TT NITF version 3.3. Tidningarnas Telegrambyrå, www.tt.se,. Yu Li et. al, 2005. A hybrid collaborative filtering method for multiple-interests and multiple-content recommendation in E-Commerce. Expert Systems with Applications 28 pp.67-77;