2.1 Classification criteria related to user model structure

The granularity of a model can be fine when an individual user is considered (individual user model) or coarse when a group of users is modeled (user group model). The composition and size of groups varies substantially from one adaptive system to another. For example, in an educational system, a team of students or a classroom may be modeled (Kosba et al., Reference Kosba, Dimitrova and Boyle2004), while in a virtual travel agency, the system may model a group of tourists that want to travel together (Ardissono et al., Reference Ardissono, Goy, Petrone, Segnan and Torasso2001). Sometimes the group includes all the users of the system. Thus, the system can provide an adaptation that enables a specific user to benefit from the collective experience (Brusilovsky et al., Reference Brusilovsky and Eklund2004).

As for the internal representation of the user, it is possible to use either stereotypes or an overlay model. Stereotypes classify users so that certain features of a new user can be assumed knowing the class or stereotype to which the user belongs. The stereotype approach, commonly used in everyday life, is also very frequently used in adaptive systems (e.g. in the Grundy book recommendation system (Rich, Reference Ritter and Young1999)). Stereotypes are simply collections of features or facets that describe a type of user. Among others, social, cultural, or knowledge stereotypes can be distinguished. For example, ‘beginner’, ‘apprentice’, and ‘expert’ are knowledge stereotypes often used in educational systems. The overlay model uses a more specific user representation which is normally based on the conceptual structure of the application domain. A well-known example is the AHAM model (De Bra et al., Reference De Buenaga, Maa, Daz and Gervs1999). However, overlay models are difficult to initialize, so they are often combined with stereotypes, producing hybrid models such as the one used in the WHURLE learning environment (Zakaria et al., Reference Vogiatzis, Tzanavari, Retalis, Avgeriou and Papasalouros2002). In addition, there are other models of representation such as the differential model (Kay, Reference Kay2000), which is basically an overlay model on expected knowledge (domain knowledge that the user should have previously achieved); the perturbation model (Kass & Finin, Reference Kass and Finin1988), which represents users as the subset of expert’s knowledge (overlay model) plus their mal knowledge (user’s errors caused by erroneous belief or lack of knowledge) and the plan model (Kobsa, Reference Kobsa1993), which consists of a sequence of users’ actions to achieve their desires or goals.

The main trace establishes the main aspect of the user that will be taken into account to perform future adaptations. Sometimes it can be difficult to identify only one trace as the main one, so several main traces can be chosen. The traces determine the content and sometimes also the structure of the UM. For example, in IR systems, the system traces the users interests and preferences about the information resources, products, or services being offered, while in educational systems, the system mainly traces the users knowledge about the concepts or information resources that must be learned. In fact, the issue of knowledge tracing has been extensively studied (Beck & Chang, Reference Beck and Chang2007). In addition, it is also possible to trace the location of the users or the tasks that they are performing. An example of the latter is the task-oriented web UM proposed by Jin et al. (Reference Jin, Zhou and Mobasher2005) to capture users navigational process.

The internal structure of the information stored in the UM is very different from one adaptive system to another and obviously will be influenced by the three previous criteria. For this reason, distinct approaches to structuring an UM are found in the scientific literature. The simplest organize information through a set of attribute-value pairs (which can also be triples or n-tuples), which can be adapted easily to a table format. An example of this is the UM defined in AHAM (De Bra et al., Reference De Buenaga, Maa, Daz and Gervs1999). Others alternatives are formal approaches such as Bayesian networks (Henze & Nejdl, Reference Henze and Nejdl1999) or Petri nets (Jianghui, Reference Jianghui2006) and the conceptual approaches which will be discussed later. A more structured approach is the episodic approach (Weber, Reference Villon and Lisetti1996), in which information is organized as a sequence of episodes that in turn are structured. This alternative permits combining several ways of structuring. An example is similarity-usage based retrieval (SUBR) (Khemani et al., Reference Khemani, Sadiq, Bangani and Rao2006), a recommender system that can learn from its experience with the user. The UM in SUBR is an episodic memory of the user’s preferences represented by triplets of the form (description, lesson, usage).

As for the general nature of the information stored in the model, it is possible to set up two types of UM (Rich, Reference Ritter and Young1999): short term if the UM contains very specific information (such as the last sentence that the user has typed), and long term if it stores more general information (such as the users experience level in mathematics). Usually, the more specific information makes it possible to design local or short-term adaptation strategies, and the general information permits the development of global or long-term strategies. Most UMs are hybrid in this regard because they combine both types of information.

Conceptual models manage an explicit representation of the concepts and conceptual relations associated with the information managed by the system. Most conceptual models use ontologies to structure the UM. This is the case with the semantic web usage log preparation model (SWULPM) (Zhang et al., Reference Weber2007), which represents the UM by a concept graph called a c-graph. Among the models based on ontologies, it is possible to distinguish those that use a general ontology (e.g. GUMO (Heckmann et al., Reference Heckmann, Schwarzkopf, Mori, Dengler and Kröner2007)) and those that use a domain-dependent ontology. The authors believe that knowledge-domain-dependent ontologies are usually more suitable for educational applications. This is why, in their proposal for creating educational hypermedia systems, the SEM-HP model (Medina-Medina et al., Reference Medina-Medina, Molina-Ortiz and García-Cabrera2005), the UM is built on the conceptual structure underlying the information domain provided to the user.

Cognitive models store information related to the user’s cognitive resources, such as perception, memory, or processing. The aim of using a cognitive model is usually explaining and predicting the users behavior. An example is the HTAmap framework (Heinath et al., Reference Heinath, Dzaack, Wiesner and Urbas2007), which provides a cognitive model based on predefined activity patterns or the work presented in Ritter and Young (Reference Ryan and Deci2001) to improve the design of interfaces. Cognitive models usually rely on psychological theories and can focus on a specific task, a cognitive process, or on the interaction between two or more processes. In IR systems, they are usually used to reveal the search strategies followed by the users of the system, while in educational systems, they are focused on the learning process. There are also psychological UMs that are able to discover the emotions of the user. For example, the Psycho Physiological Emotional Map (Villon & Lisetti, Reference Van Labeke, Brna and Morales2006) takes psychological and physiological measures and establishes associations between them to define the users emotional map.

To a greater or lesser extent, all adaptive systems must perform a monitoring process to supervise the users activity. Depending on which aspects of the user can be supervised, the information stored in the model and consequently the main trace changes substantially. For example, one can track user interface events, location, movement, audio, video, etc. In information systems, counting the number of accesses to a document and the time spent can be enough, while in life-logging systems, the monitoring is much more complex and requires the use of several devices, such as cameras, microphones, global positioning system (GPS), accelerometers, and gyroscopes. For example, the system proposed by Aizawa et al. (Reference Aizawa, Tancharoen, Kawasaki and Yamasaki2004) enables recording of all the users experiences during the day: places visited, conversations that have taken place, documents handled, and so forth.

Finally, an UM can be context dependent or context independent depending on whether contextual information is considered. There are many different types of contexts, for example, the technological context (features of the technical equipment used by the user) or the institutional context (role of the user in the organization where he or she belongs). An instance of a contextual UM can be seen in the Home Amigo intelligent environment (Vildjiounaite & Kallio, Reference Vildjiounaite and Kallio2007), which stores the users interaction history in tuples of the form (context, user action). This system, among its other functionalities, adapts the TV programming depending on who is in the room and configures the shopping list taking into account forthcoming holidays and trips.

2.2 Classification criteria related to user model management

Once the UM structure has been defined, two key aspects are its initialization and update. The initialization can be total, partial, or null, depending on whether, when the model is created for a new user, the system defines all, some, or none of the information. This initialization can be performed by default or using registration forms.

Regarding update, it is said that an UM is static if the information stored in it is rarely or never updated, while an UM is dynamic if it frequently changes adjusting its information to the users evolution. For example, in an educational UM, the learner model must evolve in parallel with the progress achieved by the student in the learning process. This update can be explicit or implicit. In an explicit or active update, the user is requested in some way to collaborate. The most typical example involves update forms (tests and surveys) that must be filled in by the user. An implicit or non-intrusive update is performed without the user being interrupted in any way while using the system. The most typical example is when the UM is updated only by observing the users behavior. This is the case in the PEACH museum (Stock, Reference Terry2007), where, among other non-intrusive data, the sequence of visits and the time spent are taken into account to guide the user within the museum.

The systems that perform an implicit update of the UM carry out an inference process to deduce the necessary changes in the model from the user’s behavior. Going one step further, some systems are able to predict the users behavior and anticipate certain aspects of the model such as the users goals, actions, and preferences. Among predictive models, two main approaches are used (Zukerman & Albrecht, Reference Zakaria, Moore, Ashman, Stewart and Brailsford2001) as follows: content-based and collaborative models. Content-based models are used when past behavior is considered to be a reliable indicator of future behavior, because the approach used here is to build the UM using data about the user’s previous behavior. Collaborative models are used when it is possible to assume that a user will behave similarly to others. In this case, the UM is built using data about a group of users, although it is used to make predictions about an individual user. For example, the navigational sequence performed by a group of students in their learning process can provide guided routes for new users. These two approaches have often been combined, giving rise to hybrid systems that perform a content-based prediction based on the result of a collaborative prediction or vice versa. For example, an educational system could predict how much time a student will take to master a topic, taking into account the student’s previous learning process and the time that other students in the same class have taken to learn that topic.

Closely related to predictive models are statistical models. A statistical UM uses sample results to make statements about an unknown parameter related to the user. Each parameter can represent an aspect of the future behavior of the user, such as goals, preferences, problems, actions, or location. In the area of artificial intelligence, many techniques exist that can be applied to statistical models (Zukerman & Albrecht, Reference Zakaria, Moore, Ashman, Stewart and Brailsford2001) such as: linear models, term frequency, inverse document frequency (TFIDF) -based models, Markov models, neural networks, classification, rule induction, Bayesian networks, and many others.

No matter how the UM is updated, with regard to the stability of the information, a distinction is made between a long-term UM and a short-term UM. In a long-term UM, the model is retained across the different working sessions that the user participates with the system. Therefore, the model’s life is often very prolonged in time, and its contents can be used to define different states in the user’s evolution. In a short-term UM, the temporal validity of the UM is very limited, and it may even be restricted to one session. This is the case in the medical-record management system presented by Ferri (Reference Ferri1996), which manages a short-term individual model which is created in each session and deleted at the end of it to personalize each medical consultation. In the case of educational systems, the UM is usually hybrid, because it needs to manage both relatively permanent information, such as the level of the student in each subject, and other, more transient information such as documents visited or exercises solved.

According to Jameson (Reference Jameson2001), there are mainly two types of inference methods for the learning and decision-making necessary for adaptation: theory-based and data-based. Theory-based inference methods work with predefined models which are then adapted to each individual user. Often these methods use stereotypes to construct theoretical models. In a data-based inference method, the main source of information processed to perform the adaptation is a set of system usage data obtained during user interactions.

In the second case, user modeling is particularly important, and it can be labeled as push or pull depending on the direction of the inference (De Vrieze et al., Reference De Vrieze, Van Bommel, Van Der Weide and Klok2005). In a push UM, the inference is done when pushing the data to the UM. Therefore, the model stores final values, that is, the information deduced through the inference process. These models are usually well-structured, easily understood, and of a relatively small size, and use update rules to propagate user events into the UM values. Therefore, when performing the adaptation, the model is queried without further processing. Push models are very popular in educational systems, like, for example, the AHAM model (De Bra et al., Reference De Buenaga, Maa, Daz and Gervs1999). Unlike push UMs, in a pull UM, the inference is done when pulling the data from the UM. Hence, a pull UM contains all the events performed by the user. In other words, the model is a log of the user’s history that will be processed at the moment when the adaptation is required by deriving high-level attributes from the stored events. Models of this type are more expandable, but have disadvantages regarding size and speed. Of course, there are hybrid UMs in which the inference is done in both directions.

Similarly, it can be said that UM usage regarding the direction of the adaptation can be of two types: exploitation, when the adaptation tends to make the UM converge, and exploration, when the adaptation tends to diverge from the UM. Hence, exploitation of the UM has a conservative nature and involves performing an adaptation process in the same direction as the information stored in the UM. An example, in an educational system, would be to recommend to the student only the kind of documents that completely fit the preferences and interests captured in the model. On the other hand, exploration emphasizes an adaptation that slightly deviates from the information stored in the model. Therefore, the system gives recommendations about which it is less certain, with the aim of discovering the potential interests of the user. This approach can avoid the situation in which the system stagnates in an outdated representation of the user, as might happen with exploitation (e.g. the system recommends a certain kind of document because it is in the UM, which causes the user to visit this kind of document more often, which in turn reinforces this preference in the UM, avoiding visits to other kinds of documents in which the user might also be interested). Again, there are hybrid systems that combine exploitation and exploration, such as the personalized IR system described in Sia et al. (Reference Sleeman and Brown2007). In the case of educational systems, exploitation enables directed learning, in which the contents are presented to the students in their final form and the pupil is required only to acquire and retain them, while exploration requires the students to discover for themselves the important concepts and conceptual relationships.

UM visibility refers to the ability of the users to see the contents of their model so that they are conscious of how they are seen by the system. Regarding visibility, and depending on the decisions of privacy, two types of UMs can be distinguished: open and closed. An open UM, metaphorically called a ‘glass box’, provides a way for the user to visualize the information stored in the model. In an educational system, an open UM can motivate the user because it enables the learning progress to be explicitly observed. Some systems go a step further and allow the users to modify the contents of their UM if they believe that it does not correspond to reality. An example in which this kind of control is given to the user is the e-learning system LeActiveMath (Van Labeke et al., Reference Martin2007). On the other hand, a closed UM, metaphorically called a ‘black box’, does not allow users to see directly or to change the contents of their UMs.

As for portability, it is said that an UM is portable if it can be completely or partially reused, with small or moderate adaptations, in applications that are different from the one for which it was created. Standardization of the model is important because it allows several adaptive systems to interoperate so that users can transport their personal models from one system to another, keeping the same image in all of them. Specifically, Vogiatzis et al. (Reference Vildjiounaite and Kallio2004) propose that for standardization of UMs, the following standards should be considered: IEEE Personal and Private Information, the PAPI standard, the IMS Learner Information Package, and the LIP standard. Another important issue related to portability is the privacy of user data, that is, the user must be informed and agree that data recorded on a system to be exported and used in a different system.

With regard to centralization, most UMs are centralized, but there are also decentralized UMs that allow information to be stored in distributed repositories. This is the case for the decentralized UM architecture proposed by Heckmann et al. (Reference Heckmann, Schwarzkopf, Mori, Dengler and Kröner2007), which structures the information by sentences written in UserML (a markup language for UMs) stored in distributed repositories using extensible markup language (XML), resource description framework (RDF), or structured query language (SQL).

3.1 User modeling in SEM-HP

One of the main aspires of our taxonomy is to provide a systematic mechanism to describe an adaptive system from the multiple perspectives of the UM managed to produce the adaptation. In this spirit, the authors are going to describe, via the taxonomy, their own adaptive system, the SEM-HP model (Medina-Medina et al., Reference Medina-Medina, Molina-Ortiz and García-Cabrera2005). This is a semantic, systemic, and evolutionary model to design AHSs, which proposes an incremental design process and a specific architecture based on this process. The architecture is divided in four interrelated subsystems. The memorization subsystem is responsible for maintaining the information provided by the hypermedia system. This information is organized using a conceptual map (represented by a semantic net), which shows the underlying concepts to each information item and the relations between them. The authors name $$CS_{M} $$ (Conceptual Structure of Memorization) to this conceptual map. The presentation subsystem permits to create partial views of the complete conceptual map with the intention of focusing on more specific knowledge sub-domains. Each partial conceptual map is denominated $$CS_{P} $$ (Conceptual Structure of Presentation) and contains a subset of the concepts, information items, and relations included in the $$CS_{M} $$ . The navigation subsystem establishes the navigability of the conceptual relations included in the different views. Finally, the learning subsystem is responsible for making adjustments to the hypermedia system during the browsing of the user.

Subsequently, the authors use the taxonomy to describe the learning subsystem of SEM-HP in terms of adaptation and user modeling. Figure 3 shows a possible representation of the taxonomic description performed for the SEM-HP’s UM. This description also could have been visualized using others graphical formalism, for example: templates of specification (such as Tables 2 and 3), trees, graphs, or markup languages such as XML.

Figure 3

Learning Subsystem of SEM-HP

Table 2

User model for education: structure

Table 3

User model for education: management

From the perspective of the structure, Figure 3 indicates that in SEM-HP two UMs with different levels of granularity are maintained: an individual UM for each user in the system and a group UM for the set of all the users. Both models are structured in episodes. In an individual UM there is an episode for each CS _M browsed for the modeled user, while that in the group UM there is an episode for each CS _P in the system. Each individual episode is represented using an overlay model on the concepts and information items in the CS _M associated with the episode. The overlay model is structured as a set of pairs (attribute value) and stores for each element (item or concept) the degree of knowledge that the user has about that element. So, the user knowledge is the main trace of the UM, although their preferences, interests, and goals also are stored. In addition, the system monitors the information that the user visits to register the number of times that each information item is accessed. Finally, a few general features (such as the user experience in hypermedia navigation) are represented using stereotypes (‘novice’, ‘expert’, etc.). Each episode in the group UM is organized using a transition matrix. The transition matrix stores the number of times that some user has followed a conceptual relation. That is, given any two concepts, A and B, in the CS _P, the associated matrix states the times a user had visited an information linked to the concept B when the last information he had visited was linked to the concept A.

From the perspective of the management, Figure 3 specifies that in SEM-HP the initialization of the UM is hybrid: personal data require explicit initialization but the rest of the data have a default initialization (e.g. the transition matrixes are created with zero value for all the conceptual relations). The update of the UM is dynamic and allows explicit and implicit updates. The explicit updates are possible because the individual UM is open to view and modify. The implicit updates are mainly based on the monitoring of the user’s browsing process (the inference method is data based). The inferences are performed using a push model (e.g. the degree of knowledge of the user about a item is calculated after each visit of the user by applying a set of update and knowledge rules predefined by the author of the hypermedia system) (Medina-Medina et al., Reference Medina-Medina, Molina-Ortiz and García-Cabrera2005). The stability of the group UM is short term, while the stability of the individual UM is hybrid (some attributes change frequently as the number of visits and others are more long term as the subject experience) and depends on each individual. The transition matrixes are statistically analyzed to identify changes needed in the navigation structures (e.g. unnecessary or missing elements in the CS _P) (Medina-Medina et al., Reference Méndez and Greenberg2006). In addition, the group UM permits a collaborative prediction, that is, the navigation strategies carried out by the users of the system are used to predict and improve the navigation of future users (e.g. if most of users follow a conceptual relation that does not exist in the CS _P, the relation is added and made available to new users). Finally, the UM is exploited and explored to generate the adaptation of the system during a guided navigation mode and during a free navigation mode, respectively (Medina-Medina et al., Reference Medina-Medina, Molina-Ortiz and García-Cabrera2005).

3.2 User modeling in educational adaptive systems

The adjustments made to a system are determined by the UM that the system operates, and vice versa. Therefore, user modeling is a crucial aspect of an adaptive system that must be designed with care. Throughout this article, the focus has been on adaptation in an educational environment. For this reason, before concluding this discussion, it is necessary to determine what options among those outlined above are more beneficial in this type of system. Tables 2 and 3 (based on Table 1) summarizes the set of features that the authors believe should be present in the UM of an educational adaptive system. These desirable features are organized according to the taxonomic criteria described earlier. Some of these features have already been mentioned in the preceding paragraphs, while others are mentioned here for the first time in this paper. In either case, the authors’ analysis (summarized in the tables), does not mean that all educational systems have to present exactly this set of structural and functional characteristics during their user modeling. This is just a starting point, and recommended features must be adjusted and qualified according to the particularities of each system and its environment.

As commented at the beginning of this section, many of the choices for educational UMs summarized in Tables 2 and 3 were already discussed when the taxonomy was described in Sections 2.1 and 2.2. To complete this argument, some additional considerations are added below:

• ∙ The work of Méndez and Greenberg (Reference Nicholas and Martin2006), like others in the area, show that ontologies are useful for creating models of teaching. They define an axis along which students can mentally organize their vision of the information domain underlying the knowledge domain. Ontologies also provide a meta-level of knowledge representation which permits connections between different knowledge domains. According to these studies, the authors consider the integration of conceptual models a factor that increases the learning through the use of educational systems.

• ∙ Numerous studies support the benefits of computer-supported collaborative learning (Jong et al., Reference Jong, Chan, Wu and Lin2006) in the educational process, advocating that students learn best from interacting with other group members. For this reason, the authors consider it is important to model, not only individual students, but also groups of students who work together in group learning tasks (collaborative works).

• ∙ Learning is a process that can be decomposed into episodes. Each learning episode reflects a change in student knowledge motivated by some event (Terry, Reference Terry2006). The episodes are stored and organized in episodic memory for later retrieval as part of an explicit learning process. Therefore, the authors believe it is desirable that an educational UM try to simulate episodic memory, thus establishing a second level or tier in the structure of the student model.

• ∙ The impact of emotions and other psychological factors in the performance of any task is clear. In the case of educational activities, numerous studies have attempted to answer certain questions related to psychological factors involved in learning tasks (Efklides & Volet, Reference Efklides and Volet2005). Some of these questions, mentioned in Pekrun (Reference Pogacnik, Tasic, Meza and Kosir2005), are as follows: ‘Which emotions are experienced by students in learning, and what is their phenomenology?’, ‘What is the functional importance of emotions for students’ interest, metacognition, problem solving, and performance?’, and finally, ‘Can desirable student emotions be fostered by modifying instruction and teacher behavior?’. The answers to these questions suggest that it is interesting to include in the student model parameters to represent some of the cognitive aspects that may influence the educational process.

• ∙ Different researches argue that the learning context determines the way that students address an educational process. Activity theory (Bertelsen & Bodker, Reference Bertelsen and Bodker2003) has been used to analyze the influence of the learning context. In particular, Govender (Reference Govender2009) defines three types of learning context: (a) the institutional context, divided into two main groups: face-to-face educational institutions and distance-learning institutions; (b) the disciplinary context, where given an institutional context, there will be a diversity of disciplinary contexts, each one with features that are specific to that discipline; and (c) the personal context, that is, the attitudes and objectives expressed by the students in connection with the object of study. The authors believe that all these contexts should be considered in the student model.

• ∙ Finally, the theory of self-determination (Ryan & Deci, Reference Sia, Zhu, Chi, Hino and Tseng2004) supports the statement that feelings of autonomy and competence have a highly positive impact on students during their learning. Self-determination theory is a macro-theory that deals with human motivation and personality. The theory studies the motivation behind the decisions that people make without any outside influence, focusing on the behavior of an individual which is due to self-motivation and self-determination. This theory points in the direction of giving students some control over their educational process, and therefore towards student models that reflect this.

In conclusion, as shown in Tables 2 and 3, the best feature for each criterion in an educational environment is usually the most versatile, that is hybrid solutions. In addition, solutions based on knowledge representation and those oriented to facilitate adaptation of the instructional process are preferred. Thus, in e-learning solutions is recommended to represent the conceptual domain to learn and the student’s behavior in terms of knowledge and achievements, as well as the inclusion of contextual and emotional aspects that can contribute to the acceleration and improvement of the learning progress.