Presentation is loading. Please wait.

Presentation is loading. Please wait.

Personalized Hypermedia Presentation Techniques for Improving Online Customer Relationships Alfred Kobsa, Jurgen Koenemann and Wolfgang Pohl Presented.

Similar presentations


Presentation on theme: "Personalized Hypermedia Presentation Techniques for Improving Online Customer Relationships Alfred Kobsa, Jurgen Koenemann and Wolfgang Pohl Presented."— Presentation transcript:

1 Personalized Hypermedia Presentation Techniques for Improving Online Customer Relationships Alfred Kobsa, Jurgen Koenemann and Wolfgang Pohl Presented by Lei Zan, Amy Henckel

2 Outline Why personalized systems (an example) What input to personalized systems How to acquire data How to represent and infer How to produce adaptation Conclusions

3 Introduction Personalization, micro-marketing, one-to-one marketing Provide values to customers by serving them as individuals Improve customer relationship, turn web visitors to customers Web provides a platform to realize this business model It facilitates large amount of data collection It supports dynamic creation of content/presentation It enables global presence

4 Introduction Personalized hypermedia application Adapts the content, structure, and/or presentation of the networked hypermedia objects to Each individual user’s characteristics, usage behaviour, and/or usage environment Adaptability and adaptivity Adaptability: the user is in control of adaptation steps Adaptivity: the system performs all adaptation steps automatically Adaptability and adaptivity coexist

5 Introduction Personalization process includes Acquisition Identify info. about user characteristics, usage behaviour and environment Make this info. accessible to adaptation component Construct user/usage/environment model Representation and secondary inference Express content of user/usage models appropriately Draw further assumptions about users, their behaviour & environment Production Generate adaptation, given a user/usage/environment model

6 One example: AVANTI Background A project (1996-1998) funded by the European Commission Tourist information system: assist travel planning, e.g. Transportation, accommodation, day-to-day activities Adaptation is applied at both user interface, content level

7 One example: AVANTI Demonstration Scenario: You are a student in Roma who studies history of art decides to go to Siena for one week to study the culture there. You are suggested to use AVANTI system to get information for your trip

8 One example: AVANTI You enter the welcome page and login, in order to allow system recognize you.

9 One example: AVANTI You are asked to fill in a questionnaire to get information to tailer to your specific need.

10 One example: AVANTI The system load a new page. For new users, a dialog box informs that the page has been loaded to avoid confusion.

11 One example: AVANTI Your first question is how to reach Siena from Roma. You find train route from Roma to Siena.

12 One example: AVANTI If you are interested in churches, you are presented a list of churches by selecting appropriate options.

13 One example: AVANTI A result of adaptivity: after picking one church, check route and working hours, etc, the system recognize you are interested in churches and list other church’ info as options for you.

14 One example: AVANTI Three months later, you decide to go back to Siena again. In the meantime, you have attended a course to learn how to use a computer. Moreover, you have used many other times the AVANTI system.

15 One example: AVANTI You log in and the system remembers you and welcome you in Siena AGAIN.

16 One example: AVANTI Interface Adaptivity: a list of links in the left side; no feedback dialog box; you are considered as an expert user now.

17 One example: AVANTI A result of adaptivity: shortcuts and additional navigation support for quick access are provided, as you are recognized as expert.

18 Outline Why personalized systems What input to personalized systems How to acquire data How to represent and infer How to produce adaptation Conclusions & discussions

19 What are inputs to personalized systems User data Info. about user characteristics Usage data User’s interactive behaviour Environment data (of user) Software Hardware Physical environment

20 What are inputs to personalized systems User data Demographic data Record data (e.g. name, address, phone numbers) Geographic data (e.g. area code, city, state) User characteristics (e.g. age, sex, education) Registration for information offerings Note: today’s personalized system contains mainly those demographic data and purchase data. It has high value when combined with high-quality statistical data, e.g. purchase behaviour of different user groups

21 What are inputs to personalized systems User data User knowledge (about concepts, relationships between concepts in an application domain) e.g. Generate expertise-dependent product description User skills and capabilities e.g. Adaptive help messages for UNIX commands e.g. AVANTI takes the needs of disabled people (wheel-chaired, vision-impaired)

22 What are inputs to personalized systems User data User interests and preferences e.g. Sell cars to different customers emphasizing different attributes (speed, safety, etc) User goals and plans Find information on a certain topic, or shop for some products Support users to achieve their goals e.g. Present to users only information relevant to their goals

23 What are inputs to personalized systems Usage data: interaction behaviour Observable data Selective actions Indicator of user’s interest, or unfamiliarity, or preferences Viewing time Potential indicator of user interest Ratings Indicate how relevant or interesting the object is e.g. eBay, Amazon Purchases and purchase-related actions Strong indicator of user interest

24 What are inputs to personalized systems Usage data Usage regularities: further processing of data Usage frequency e.g. AVANTI monitors how often individual users visit certain pages and introduces shortcut links Situation-action correlations e.g. Email assistant: suggest how to deal with incoming emails, based on statistics of correlations between previous emails (situations) and how user processed them (actions) Action sequences Used to recommend macros for frequently used action sequences, predict future actions

25 What are inputs to personalized systems Environment data: impact web usage Software environment Brower version and platform, availability of plug-ins, java and javascripts Hardware environment Bandwidth, processing speed, display devices, input devices locale Users’ location, characteristics of locale (e.g. noise level )

26 Outline Why personalized systems What input to personalized systems How to acquire data How to represent and infer How to produce adaptation Conclusions & discussions

27 How to acquire data User model Collection of explicit assumptions about user data Usage model Construct aggregated information about a user’s interactive behaviour from observations Environment model

28 How to acquire data User model acquisition methods Active acquisition: User-supplied information Questionnaires, initial interviews e.g. AVANTI welcome page asks questions (computers, AVANTI systems, about disabilities) Downside: paradox of the active user User wants to get started immediately and get work done soon Time is saved in the long term by taking initial time to optimize system

29 How to acquire data User model acquisition methods Passive acquisition Acquisition rules Refer to observed user actions or straightforward interpretation of user behaviour e.g. a classic domain-independent rule: “If the user wants to know X, then the user does not know X” Plan recognition Recognize user’s goal from observed user interactions Suitable for applications with a small number of goals and ways to achieve the goals

30 How to acquire data User model acquisition methods Passive acquisition Stereotype reasoning Categorize and associate a stereotype with each category Stereotype contains standard assumptions about members of that category and activation conditions Evaluate activation conditions, apply content of stereotype as assumptions to the particular user e.g. if the user is interested in childcare, activate “parent” stereotype

31 How to acquire data Usage model acquisition methods Simple technique Record user actions in order to obtain information about user behaviour Learning algorithms Memory-based learning, reinforcement learning, induction of decision tree e.g. learn situation-action correlations; these data are used to predict user behaviour in future situations

32 How to acquire data Environment data acquisition methods Software environment: http header Hardware environment Difficult to assess e.g. AVANTI evaluates bandwidth from media download time Locale Location can be recorded in database or use GPS

33 Outline Why personalized systems What to input to personalized systems How to acquire data How to represent and infer How to produce adaptation Conclusions & discussions

34 How to represent and infer Why need representation and inference Some applications operate directly on results of user/usage/environment model Some applications need user/usage model representation and further inference Deductive reasoning (from general to specific) Inductive reasoning (from specific to general)

35 How to represent and infer Deductive reasoning (from general to specific) Logic-based representation and inference e.g. Concept formalism: form user knowledge base Shortcomings of logic-based approaches Limited ability to deal with uncertainty and with changes to the user model Representation and reasoning with uncertainty Bayesian network, evidence-based, fuzzy logic approach for probabilistic user model representation

36 How to represent and infer A concept hierarchy in animal kingdom

37 How to represent and infer Inductive reasoning (from specific to general): Learning about the users: monitor users’ interaction with system and draw general conclusions based on observations Learning is used to construct “interest profiles” Interest profiles represent a user’s interest in an object, based on an assessment of his interest in specific features of the object e.g. assumption of user interest in movies is determined by preferences about actor, director and other movie features Neural network, machine learning, nearest-neighbour algorithm, induction of decision trees, etc.

38 How to represent and infer Analogical reasoning Takes advantage of the large number of users for web- based systems. Clique-based filtering Matches a single implicit profile with profiles of similar users Takes into consideration other issues Content may not be the only aspect that determines interest Content of object may not be easy to analyze by computers User’s interest may not be based on features of objects Three steps Find similar neighbors Select a comparison group of neighbors Compute prediction based on weighted representation of selected neighbors

39 How to represent and infer Analogical reasoning Clustering user profiles Uses explicit profiles to exploit similarities between users. Unlike stereotyping: classification is not rigid. Example: The Doppelganger is a user modeling server that stores profiles of different users and forms group profiles via clustering algorithm. Differences from stereotyping are strengths of values in a group model and distance between the user profile and group profile are considered.

40 How to represent and infer Hybrid approach: User profiles as learning results A kind of “closed-loop adaptation” appears in the previous learning methods. An example of hybrid approach is feature-based filtering. Feature-based descriptions are used for CDs that a user has shown an interest in and maybe also for CDs that a user has shown no interest in. A supervised learning algorithm can be used to classify CDs “interesting” and “not interesting”. Results will varying depending on the algorithm used. Primary purpose is classification of CD descriptions and recommendations based on them.

41 How to represent and infer Hybrid approach Example: LaboUr Exploits the advantages of both explicit and implicit user profiles. Learning components (LCs) and acquisition components (ACs) can select appropriate user observations. LCs generate usage statistics and patterns. ACs support heuristic rule-based acquisition from isolated observations. Together they generate usage-related or user-related user-model contents.

42 How to represent and infer Hybrid approach Example: LaboUr Decision components (DCs) refer to explicit user models. LaboUr system can provide direct access to user model contents. It can maintain several user models. Further learning components can be used for clustering user models into user group models.

43 How to represent and infer Hybrid approach Example: LaboUr

44 Outline Why personalized systems What to input to personalized systems How to acquire data How to represent and infer How to produce adaptation Conclusions & discussions

45 How to produce adaptation Adaptation production Why should hypermedia adapt for individual users? To help the users use the site more efficiently To make the user’s experience more exciting To encourage the user to stay To help the user learn, in the case of educational hypermedia There are three types of adaptation Adaptation of content Adaptation of presentation and modality Adaptation of structure

46 How to produce adaptation Adaptation of content Methods for personalizing the content of hypermedia objects and pages in accordance with user, usage and environment data. Frequently used personalization functions of content adaptation are listed. Optional explanations: giving explanations only to users who lack the necessary background knowledge to already know them. Optional detailed information: can make a hypermedia page more interesting and improve its relevance for users.

47 How to produce adaptation Adaptation of content Frequently used personalization functions of content adaptation are listed Personalized recommendations – inform users about available offerings that they may be interested in. Theory-driven presentation: caters to users following more general didactic, rhetorical or psychological principles, often combined with restricted natural- language generation. Optional opportunistic hints: based on users’ presumed interests and on current circumstances. Ex: news flashes, pointers to special discounts, etc.

48 How to produce adaptation Techniques for content adaptation Page variants Different versions of all pages are created. It is relatively simple, but cumbersome since a new page for each variation must be created and inflexible since modifications are manually done. Fragment variants: Different versions of page fragments are created. At runtime the appropriate fragments are brought together in a static page frame. The fragment could be a paragraph of text, an image, a video, etc. More difficult than page variants since web pages must be put together at runtime. Examples: CGI and Java servlets, XML, programmable web servers.

49 How to produce adaptation Techniques for content adaptation Fragment colouring Content remains the same for all users. However some elements may be marked out. Advantage: all users can see all information. There is less of an impact if an error is made assessing the user. Disadvantage: Not much change from one user to another. Not widely used.

50 How to produce adaptation Techniques for content adaptation Adaptive stretchtext Text that the user can extend or collapse by clicking on it with the mouse. Usually a few words or one sentence. Based on the user model, can be automatically expanded or collapsed by the system. Advantage: user can further adapt the page if it was adapted appropriately by the system. This information can be kept to refine the user model. Adaptive natural-language generation Uses alternative text descriptions for different users depending on their level of expertise. Can be used with stretchtext to make it more meaningful for the user.

51 How to produce adaptation Adaptation of presentation and modality Describes methods for changing the presentation and media format and the interaction elements. The information content stays the same. The format and layout changes. Adaptations for multimedia presentations are often based on the user’s preferences. These preferences are communicated via a short interview or on a machine-readable data carrier as seen previously. System performance example Videos and large images may be replaced by a link and estimated download times.

52 How to produce adaptation Adaptation of presentation and modality Example: AVANTI Special type is a change in modality – images to text, text to audio, etc. Bases the selection of different modalities on the user’s physical abilities. Example: a map of Siena, Italy that includes museums, churches, restaurants and hotels for the blind would have an auditory output of:

53 How to produce adaptation Adaptation of structure Describes methods for personalizing the presentation of links based on user, usage and environment data. Techniques for structure adaptation Need to distinguish whether or not a link anchor is part of a context. More freedom moving non-contextual links than moving contextual links. Collateral structure adaptation Content adaptations can contain link adaptations if fragment variants contain links which are not presented to the user in one of the variants.

54 How to produce adaptation Adaptation of structure Techniques for structure adaptation Adaptive link sorting Simple technique only used for non-contextual links. Examples: Ranking target web pages based on their relevance to users’ interests and goals, and background knowledge. Linked lists in HYPERFLEX show the relevance of the links to the current page and specified goal. User can give input. Also used for presenting ranked lists of recommended items, such as movies, electronic equipment, books, etc. Other possible uses are link sorting based on frequency of use. However caution should be exercised since automatic sorting of menu items can confuse users.

55 How to produce adaptation Adaptation of structure Techniques for structure adaptation Adaptive link annotation Can be used for contextual and non-contextual links. Non-adaptive link annotation is used in all web browsers: links change colors if already visited. Adaptive hypermedia systems use different colors and symbol codes to personalize links. Examples: Using different backgrounds, colors, styles to mark recommended link anchors. An educational system use colored bullets for links to chapters or concepts based on the users current knowledge level.

56 How to produce adaptation Adaptation of structure Techniques for structure adaptation Adaptive link hiding and unhiding Adaptive link hiding means the link looks like normal text or a normal icon. Used to guide users to the most relevant pages or the most comprehensible ones to the user. Also used to hide links until the user has visited the prerequisite pages.

57 How to produce adaptation Adaptation of structure Techniques for structure adaptation Adaptive link disabling and enabling This removes the functionally of the link, but the appearance remains the same. Disadvantage: This behavior violates the principle of expectation conformance in HCI. Link disabling and enabling is currently used together with link hiding and unhiding only. Still has problems. A study showed students learned more slowly with this combination than other students, however there were positive results for learners with low prior knowledge with a system with non-recommended links disabled.

58 How to produce adaptation Adaptation of structure Techniques for structure adaptation Adaptive link removal/addition Deletes the link anchors completely, only applied to non- contextual links (unless the whole context is removed as well). Reduces the hypermedia space by removing non-relevant links. Example of link removal: Removing links to a pages that are not yet appropriate for a learner. Removing product links that are of no interest to the user. Study showed the adaptive link removal to be very effective in supporting users’ navigation by reducing the number of steps to achieve that goal. Contrasting study shows that users did not like the link removal. If a stable listing of links of used frequently, removal could violate the constancy principle of HCI.

59 How to produce adaptation Adaptation of structure Techniques for structure adaptation Adaptive link removal/addition HIPS (mobile web-based museum guide). The system shows links to nearby paintings, and distant paintings based on the users interest in topics, painters or time periods.

60 How to produce adaptation Adaptation of structure Personalization functions of structure adaptation Adaptive recommendations Most frequently used types: Recommendations concerning products: Lists of links to products and services are filtered and ranked based on user data. For example, Amazon.com, HIPS, etc. Recommendation concerning information: Lists of links to documents or other information are ranked based on user and usage data. Navigation recommendations: Links to hypermedia pages are filtered or ranked based on user, usage and environment data.

61 How to produce adaptation Adaptation of structure Personalization functions of structure adaptation Adaptive orientation and guidance Makes navigating through a site easier for the user. Personalizing a site’s overview map can mark a user’s visited or bookmarked pages. Guided site tours for first-time users is another popular method.. Personalized next buttons are used sometimes in educational hypermedia systems. It’s a very flexible method since it’s computed at run time. Results in linear navigation via direct guidance, which can be also be very effective.

62 How to produce adaptation Adaptation of structure Personalization functions of structure adaptation Personal views and spaces Bookmarking in web browsers provides a personalized access to web sites. Hierarchical lists allows the user to organize the bookmarks. One system collects URLs that were recently and most frequently used, into a short list. WebTagger enables users and groups to store, access and rate the URLs, including automatic categorization of URLs. User modeling and usage modeling could be used to realize system-supported adaptability to the users’ needs.

63 How to produce adaptation Adaptation of structure Personalization functions of structure adaptation Personal views and spaces Many portals and other websites encourage users to create personal spaces on their sites. View histories of their past actions (flights, items purchased, etc). Create lists or markers (future buys or wish lists). Define shortcuts to resources they use frequently. Specify information they want to have sent to them (news from certain categories, messages from friends (myspace.com)). Save documents and news in a personal repository.

64 Outline Why personalized systems What to input to personalized systems How to acquire data How to represent and infer How to produce adaptation Conclusions & discussions

65 Conclusions This article reviewed methods for construction of personalized hypermedia systems. The likely main applications will be in customer relationships. However personalization is not always necessarily useful, a human assistant is sometimes still needed.

66 Conclusions Application areas and their requirements. Public websites: It is more difficult to keep visitors at the site than to get visitors to the site. There are different considerations for different type of users: first-time, returning, infrequent and frequent. Information kiosks: Different situations with “walk up and use” systems by first time or infrequent users. Other factors include background noise, reduced privacy, etc

67 Conclusions Application areas and their requirements Web-capable appliances and mobile devices: These range from internet kiosks, car mounted devices, pdas to cell phones, etc. Need to consider the variety of output devices, connection speeds, capabilities, and uses. “Universal access”, “design for all”, and “user interfaces for all”: Basically software should be designed so there are no barriers with people with special needs: disabilities, elderly users and those with different cultural background. It is more economically feasible in the long run to develop more generic applications that can be adapted or can adapt themselves to people’s needs.

68 Conclusions Acquisition and representation A user should be exposed to content and not lengthy registration procedures or initial interviews. If the user interview method is chosen, it should be restricted to 2 or 3 core questions that are embedded in the main page. Because of this, adaptively has to be primarily based on user data from sources like Marketing data, stereotypes and data based on user behavior. Example: Giving information on currently owned vehicle and zip code may provide a very rough stereotype

69 Conclusions Acquisition and representation Other adaptation methods: Offer adaptivity to the user as an option. Example: users can be offered a short quiz and promised better recommendations for products and services. Users should always feel in control and have the option of correcting, undoing, or ignoring adaptive modifications. Example: they should be able to re-sort a generated list. Personalized systems should allow the user to view all items.

70 Conclusions Acquisition and representation It is recommended to log the user interaction and navigation behavior. It is not recommended to log usage data on the micro- interaction level (ie. tracking mouse movements). Amount of data is too large In the no-so-distant future, identifying and authorizing users could be based on physiological characteristics.

71 Conclusions Acquisition and representation Privacy concerns of users have to be taken very seriously. Tell users what is done and the added value of providing this information. Users should be able to opt out of logging their data and should be able to view their user model. Users could be highly anonymous.

72 Conclusions Acquisition and representation Integration with other customer-relationship applications and data User’s implicit profiles can not be directly transferred to other systems. In current customer-relationship components, a lot of data exists in explicit form (e.g. demographic marketing data). Exploiting this data would be easier with explicit models. Since both models have their advantages, it is best to use both in a hybrid system.

73 Discussion Any questions?


Download ppt "Personalized Hypermedia Presentation Techniques for Improving Online Customer Relationships Alfred Kobsa, Jurgen Koenemann and Wolfgang Pohl Presented."

Similar presentations


Ads by Google