Generic Adaptation Framework for Unifying Adaptive Web-based Systems
The Generic Adaptation Framework research project aims to develop a new reference model for the adaptive hypermedia research field. The new model will consider new developments, techniques and methodologies in the areas of adaptive hypermedia and adjacent fields (Data Mining and Machine Learning, Semantic Web, Open Corpus Adaptation, etc.).
Adaptation process Search and Adaptation Provenance and Adaptation Recommendation and Adaptation
Please check GAF publications HERE
(fig.1) GAF Conceptual Adaptation Process sequence
By Generic Adaptation Process we mean the interaction in AHS which starts with the goal statement, exploits features of the user and domain models in different contexts and adapts various aspects of the system to the user. This sequence of user-adaptation actions could be aligned according to the classification of AH methods and techniques which results in the adaptation sequence, coupling the ‘layers’ of AHS by means of Adaptation Process.
(fig.2) Search compliance with GAF structure and process
The Search Process complies with the reference structure of AHS as follows:
• The User states the goal thus formulating a new search query, which can be considered as stating or choosing a particular concept (set of concepts) to follow in AHS. It can be interpreted and aligned with DM (availability of concepts, concept structures and sequences, etc.) and UM (considering user competencies, preferences, experience, etc.) thus reformulating and refining the search query (matching it with the common lexicon or using semantically related terms);
• The Domain Model is defined by the search index, representing keywords used to facilitate fast and reliable information retrieval, which is acquired from the Resource Model (and essentially WWW). The index information is obtained from WWW by means of crawling which is similar to the process of resolving content information of a concept in AHS;
• The Context Model defines user and usage context properties such as IP address, user profile/stereotype, or search and result histories accordingly;
• The Group Model refers to maintaining a collaborative profile of the user or stereotyping search results by location or user age group and gender, which later can be used to rank and recommend results;
• Retrieving and updating UM refers to storing and accumulating UM search history which can be used to reformulate queries or retrieve personalized results;
• Application and Adaptation Models may refer to the Search Engine and Ranking mechanisms, however it may not be entirely clear how to distinguish some particular parts of those. Here we would refer to the Adaptation Model for Ranking, since they both to some extent perform adaptation of the results. The Application Model then serves as the core of the system: coupling other layers and dispatching information in AHS or performing a search as the Search Engine;
• The Presentation Model renders search results and presents a ranked result list, snippets, additional rank information, groups result, etc.
Provenance and Adaptation:
Considering the question-centric, extensive definition of
the W7 Provenance Model and the AH methods and techniques classification
questions we combine and align the questions and corresponding answers. Such
an alignment will be able to provide complementary features description.
Here we investigate commonalities and similarities in the semantics of the
answers and meanings of these questions, emphasizing the idea that
provenance information can be very useful in AHS and at the same time
provenance information can help to reason in AH, for example providing more
explanations to the end user or making the system more trustworthy. In
Table 1 we map questions and look for common
understanding in-between Provenance and AH:
“What?” — answer to this question on the one hand
describes the way domain information is represented in the system (hierarchy of
concepts, ontology, etc.) and on the other hand shows what events in the system
these data objects can affect;
“Who?” (“To Whom?”), “Which?” — answers to these questions give us an idea of the UM environment: Which? defines the device capabilities and in general Who? represents the user profile. They also describe the set of devices and agents involved in the process from the provenance point of view and can be used to select the target group of users, representing the high-level user division and defining the group adaptation parameters;
“To What?” — answer narrows down the user profile to a particular set of attributes involved in the adaptation process (accessed and updated by the system to retrieve or refresh the current state of the user knowledge, interest, competence, etc.). These are usually domain dependent attributes. There is no actual match on the provenance question here, however the history of UM attributes’ access and updates directly refers to storing and harvesting provenance data from user logs;
“Why?” — answer determines the set (one-at-a-time or a sequence) of goals of adaptation and describes the set of reasons for initiating the concerned adaptation process. Thus, these two indicate the premises of the adaptation process in general, provide arguments and describe the way adaptation is initiated;
“When?” and “Where?” — answers are registered as a part of the provenance model events. The AHS keeps track of these changes and interprets this data to be used as the input for the reasoning component, which should take into account this time and place contextual information;
“How?” — answer provenance data records event-action sequences, describing mostly the syntax of these changes, on the other hand AHS describes the semantics (understanding of these cause-event relationships), contributing to the picture of the reasoning model. As a whole it describes AE functionality of the system;
|questions||Adaptive Hypermedia System||Provenance Model||comments|
|What?||Domain Model||denotes the sequence of events that affect the data object||answers describe the sequence of events when the user gets access to the domain information and acquires domain knowledge|
|Who? (To Whom?) Which?||describes the user profile selection (or/and device usage) (e.g. can be used to select a group or target users)||the set of all agents and/or devices involved in the process|
|To What?||UM attributes (selecting particular attributes that are accessed and updated within the concerned adaptation process)||no actual representation in terms of provenance question, however historical information on accessing and updating UM represents provenance information|
|Why?||stating the adaptation goal(s) (might be a domain concept, representing either a new goal to follow or a sequence of concepts)||the set of reasons for triggering a particular event (evidence of what has happened)||reasons and goals are complementary, indicating the premises of the adaptation process|
|When? Where?||Application Model (which serves as the core of the system: coupling other layers and dispatching information in AHS) and Context information keeps track and interprets the context information||the set of event times and locations||contextual information in general|
|How?||describing AH methods and techniques on a conceptual and implementation level (Adaptive Engine (AE) functionality); explains the sequence of event-actions; describes the semantics of cause-effect relations||the set of all actions leading up to the events (keeping track of the events, and corresponding action in the system); describes the syntax of events and actions recorded;||
in pair provenance and AH describe the syntax
and semantics of AE functionality (record events
and actions and show cause-effect rela-
(Table 1) Aligning Adaptation and Provenance questions
Recommendation and Adaptation:
(fig.3) Recommender system compliance with GAF structure and process
Figure 3 presents the picture of compliance of a recommendation and an overlaying Generic Adaptation Process (GAP) ‘sequence chart’. Here we present GAP process chart constructed by coupling the layers of a general purpose AHS. We assign recommendation steps to a single layer or a transition in the system and discuss this compliance presented in the figure further in this section. Though we are facing certain issues distinguishing Recommendation Engine functionality, in particular filtering and ranking mechanisms (in this respect Application Model (AM) and Adaptation Model/Engine (AE) can be treated accordingly) we could align recommendation and describe its functionality (in terms of aforementioned models) with GAF terms. On the one hand this proves a generic property of GAF, and on the other hand it opens new horizons to facilitate and generalize recommendation aspects such as in hybrid or knowledge-based recommenders. Hereafter we are going to summarize the compliance of the recommendation process with the reference structure of AHS and explain the building blocks and interactions:
• The User states the goal thus formulating a new recommendation query thus inference over the user preferences is made (which is particularly interesting in knowledge-based recommender systems). This step can be considered as stating or choosing a particular concept (set of concepts) to follow in AHS. The goal can be interpreted and aligned with DM (availability of concepts, concept strictures and hierarchies, etc.) and UM (considering user competencies, preferences, experience, interests, etc.). The same way the recommendation query can be reformulated, refined or aligned to match with the related user preferences using semantically related terms to get better recommendations by inferring closely related items (e.g. name of a favorite film director relates to a certain movie genre); UM is significantly important in collaborative filtering and corresponding recommender systems;
• The Domain Model (DM) is defined by the knowledge structure of the domain, representing keywords and terms together with the relationships which can be used to facilitate fast and reliable information retrieval and filtering of the concerned itemspace, where the re- sources are defined by RM. On the other hand DM may represent the feature space in case of content-based recommendations where content is again retrieved from RM. It happens more often in the recommender systems that the domain is represented by an ontology in order to facilitate more elaborate reasoning over the items relationships and present more accurate recommendations;
• The Context Model defines user and usage context properties such as IP address, time, other activities of the user, etc. Modeling Context gives an possibility to consider context-aware recommendations and adaptation, both from the usage and the user point of view;
• The Group Model (GM) refers to maintaining a collaborative profile of the user(s) or stereotyping filtered results by location or user age group and gender, which later can be used to rank and recommend results for a particular user or mediate user models associated with different groups. Group Model in general may represent and serve heterogeneous user groups by looking up commonalities in profiles to form groups or similarities in the group system usage (usage patterns) to recommend next best items both in the context of recommender and adaptive system/application for the users within the same group;
• Retrieving and updating UM refers to storing and accumulating users’ rankings and recommendation history which can be used to reformulate system queries or retrieve personalized recommendations by finding similar patterns in the users’ system usage;
• Application and Adaptation Models may refer to the Recommender Engine involving Filtering and Ranking mechanisms, however it may not be entirely clear how to distinguish some particular parts of those between layers of AHS. Here we would refer to the Adaptation Model for ranking and recommendation rules, since they both to some extent perform adaptation of the results. The Application Model then serves as the core of the system: coupling other layers and dispatching information in AHS and Recommender respectively and performing a corresponding filtering method retrieving information from UM and DM for collaborative and content-based filtering respectively. Usually search and recommender engines are more robust and flexible for introducing or discovering new rules compared to the Adaptation Engines. However the rule systems which are conventionally used in AHS can easily facilitate reasoning in recommender systems (e.g. ECA type of rules to determine static recommendation filters such as gender or location aware, or at the same time serve as the basis for semantic reasoning to look up for a related concepts that the user might be interested in). AHS will also provide the so-called ‘higher-order adaptation’ capabilities. They will monitor the user’s behavior also to adapt the adaptation behavior by discovering new or refining old rules;
• The Presentation Model renders recommendation and adaptation results in such a way that a recommendation list is presented in a form of a ranked list, snippets, additional ranking information, result groups, etc. By applying AH presentation and navigation techniques here we may present not only ranked or sorted lists, but the whole new spectrum of interaction becomes available to enhance user experience in recommender systems (e.g. (de)emphasizing results in the list, summarizing results, navigating through the list, presenting contextual and non-contextual links, annotating, etc.)
E. Knutov, P. De Bra and M. Pechenizkiy. AH 12 Years Later: a Comprehensive Survey of Adaptive Hypermedia Methods and Techniques, New Review of Hypermedia and Multimedia 15(1), Taylor & Francis, UK, pp. 5-38, 2009.
E. Knutov, P. De Bra and M. Pechenizkiy. Generic Adaptation Framework: a Process-Oriented Perspective, Journal of Digital Information 12(1), USA, 2011.
E. Knutov, P. De Bra and M. Pechenizkiy.
E. Knutov, P. De Bra, D. Smits and M. Pechenizkiy. Bridging Navigation, Search and Adaptation. AH Models Evolution, WEBIST'11: Proceedings of the 7th International Conference on Web Information Systems and Technologies, SciTePress, pp. 314-321, 2011.
E. Knutov, P. De Bra and M. Pechenizkiy.
E. Knutov, P. De Bra and M. Pechenizkiy.
J. Hannon, E. Knutov, P. De Bra, M. Pechenizkiy, B. Smyth and K. McCarthy. Bridging Recommendation and Adaptation: Generic Adaptation Framework - Twittomender compliance study.
E. Knutov, P. De Bra and M. Pechenizkiy.