Adaptive hypermedia built the field's design taxonomy, then stalled on a negative result

Adaptive hypermedia produced the field's only complete taxonomy of adaptation, and then its empirical programme returned a result its authors had not promised. The argument here is that the taxonomy outlived the field that built it, while two disciplines that had made it usable did not. Authoring asks how a team specifies an adaptive behaviour for reuse. Scrutability asks whether the person the system adapts to can see the adaptation and refuse it.

The second of ten units, this one asks why the programme that produced the taxonomy stopped short of both. The third unit asks what the mechanism behind a recommendation fixes about the reason the interface can show.

A design space, not a feature

The founding move was to describe adaptation as a space rather than a list of systems. Brusilovsky (1996) split it into adaptive presentation, which changes what content a system shows and how, and adaptive navigation support, which changes the links. Naming the split turned adaptation from a feature a system has into a decision a designer makes.

Navigation was an early adaptation target. Kaplan, Fenwick and Chen (1993) built HYPERFLEX, which recommended information from a reader's goals and context using cheap updatable associative matrices. They reported that goal-based adaptive sorting reduced the number of hypertext nodes a reader had to visit. The effect on search time itself was not significant. Two qualifications matter: the studies were informal, and the authors aimed to identify when such adaptation is useful.

Weber and Specht (1997) joined both halves with ELM-ART II, a WWW-based LISP textbook. There an overlay model combined with an episodic user model drove content selection and adaptive navigation. It annotated links and chose the next curriculum step on demand. Its empirical result is the first warning: the techniques had different effects on different types of users during the first lessons.

The taxonomy settled while the practice stayed local

A decade on, researchers had organised the classification into a field but not into a shared implementation. Knutov, De Bra and Pechenizkiy (2009) surveyed twelve years of adaptive hypermedia and set out requirements for a new reference model. Many systems stayed tightly integrated with one application, or used few techniques, so the field could state no generic model.

INSPIRE shows the taxonomy in use. Papanikolaou and colleagues (2003) built a system in which a learner model carrying knowledge level and learning style drove curriculum sequencing, adaptive navigation support and adaptive presentation. The learner could move between full system control and full learner control by intervening in lesson generation.

The empirical programme returned a condition, not an effect

The empirical programme did not deliver the effect its promoters expected. The founding demonstration rested on informal studies of one system (Kaplan, Fenwick and Chen, 1993). The joined educational system reported differential effects by user type rather than a main effect (Weber and Specht, 1997). Yudelson and Brusilovsky (2008) built a service-based approach to link annotation, separating decision-making from the application into a personalisation service. They tested feasibility by a performance evaluation. The question asked was whether the service runs, not whether learners improve, while the techniques migrated into adaptive Web applications (Brusilovsky, 2007).

The result is negative in a precise sense: link adaptation helped under conditions, and the conditions were never assembled into a general design rule.

The field built scrutability into a working system in 1996

Hook and colleagues (1996) built the POP adaptive hypertext help system on one argument. Careless adaptation produces an unpredictable, obscure and uncontrollable interface, so a design must let users inspect the adaptivity mechanisms and control the results. INSPIRE carries the commitment into teaching, because a learner who can take over lesson generation can inspect and overrule the adaptation (Papanikolaou et al., 2003).

Hook and colleagues supply the working version. The design expresses the adaptivity mechanism in the domain concepts the reader already uses, rather than in the system's own model. A reader can then inspect what the system inferred and control the result. The design followed from empirical findings about adaptive interfaces turning unpredictable, obscure and uncontrollable, and the paper reports both the design considerations and those findings. It is the earliest case here of scrutability built into a system rather than argued for.

The successor programme measured the model rather than either discipline. Xie and colleagues (2019) reviewed journal articles from 2007 to 2017. They found that personal data such as preferences, achievements, profiles and learning logs had become the main parameters of adaptation. Most systems still supported only traditional computers and devices, so the learners the review reached were those with a desktop and a stable connection, not a reader on a shared phone. Those dimensions describe the model and the study, not the authoring and scrutability constraints.

The claim that the successor field dropped the discipline is this unit's argument, not a measured result. The papers establish that the requirement existed and that a system met it.

Authoring was the unsolved problem, and every fix stayed local

A taxonomy is useful only if someone can build inside it. De Bra and colleagues (2003) built AHA! as a general adaptive hypermedia architecture. It grew from conditional explanations and link hiding for one online course into a platform whose authoring tools added adaptive features to courses, museum sites and encyclopaedias. Stash, Cristea and De Bra (2004) showed the price of authoring learning-style adaptation. An author must express a pedagogy in a form the engine can execute. Brusilovsky (2004) stated the reuse problem as architecture in KnowledgeTree, and distributed reusable learning activities so adaptation could be shared at teacher level. Henze and Nejdl (2002) took that logic to open corpora and estimated knowledge from standard metadata and ontology subgraphs.

The pattern is a field re-deriving the same fix, reuse through architecture, with each fix remaining local. Authoring discipline is more than tooling. It requires an author to specify an adaptive behaviour in a form another person can reuse, and so it applies scrutability to the author rather than the user.

The contested question: was the stall a failure of technique or of evidence?

Two positions exist, and they disagree about what the field's results mean.

The first holds that link adaptation did not pay, so the move to statistical recommendation was rational. The founding benefit came from informal studies (Kaplan, Fenwick and Chen, 1993). The joined educational system reported effects that differed by user type rather than a general gain (Weber and Specht, 1997). Evaluation had become an architecture performance test (Yudelson and Brusilovsky, 2008). Twelve years in, the field could state no generic model, because systems remained bound to single applications (Knutov, De Bra and Pechenizkiy, 2009). On this reading the taxonomy was complete and the effect was not.

The second holds that the effect was real but conditional, and that the field failed to keep the disciplines that would have made the conditions reachable. Kaplan and colleagues reported that HYPERFLEX reduced the number of nodes a reader had to visit (Kaplan, Fenwick and Chen, 1993). The study set out to find when such adaptation is useful. Scrutability had a working design a decade before it was needed (Hook et al., 1996). The field had recognised the reuse problem and partly solved it (Henze and Nejdl, 2002; De Bra et al., 2003; Brusilovsky, 2004). On this reading the result says as much about an evaluation practice that measured short-term effects as about the technique.

What would settle it is a comparison this unit's literature does not contain. The decisive study would hold the authoring budget constant and adapt links at different strengths against a stable baseline. It would measure an outcome beyond short-term achievement and report the quality of the model driving the adaptation. A reader could then attribute a null result to the technique, the authoring or the measure. A cheaper test would re-run a link-adaptation study at modern scale.

The two positions rest on different kinds of evidence. The first rests on a founding demonstration its own authors call informal (Kaplan, Fenwick and Chen, 1993). It adds one educational system whose study reports effects that differ by user type (Weber and Specht, 1997). It adds an evaluation of one architecture's performance (Yudelson and Brusilovsky, 2008). The second rests on the same results read the other way. It adds one help system read with its own users (Hook et al., 1996). While the dispute is open, hold the authoring budget and the measure constant and report both. The strongest objection is that the inheritance claim is this unit's argument, not a measured result.

Where this is heading

The vocabulary has outlived the systems. KnowledgeTree (Brusilovsky, 2004) with the open-corpus modelling of Henze and Nejdl (2002) anticipated personalisation delivered as a service rather than as a rule inside one application. The service-based approach of Yudelson and Brusilovsky (2008) is that idea tested. The outstanding problem is the one Knutov, De Bra and Pechenizkiy (2009) named. The successor programme reviewed by Xie and colleagues (2019) has not taken it up. Whether the taxonomy can be re-inherited with its constraints is the open question. Hook and colleagues (1996) answered it in 1996 for one help system. Whether the field has answered it at scale is not something these papers can settle.

Sources

• Brusilovsky, P. (1996). Methods and techniques of adaptive hypermedia. User Modelling and User-Adapted Interaction. doi:10.1007/BF00143964 Reviewed adaptive hypermedia and split adaptation into methods that change content and presentation and methods that change the links. It is this unit's founding taxonomy.
• Brusilovsky, P. (2004). KnowledgeTree: a distributed architecture for adaptive e-learning. Proceedings of the 13th international World Wide Web conference, Alternate track. doi:10.1145/1013367.1013386 Presented KnowledgeTree, an architecture built on distributed reusable intelligent learning activities, aimed at bridging learning management systems and adaptive hypermedia. Its targets were component-based assembly and teacher-level reusability.
• Brusilovsky, P. (2007). Adaptive navigation support. In The Adaptive Web: Adaptation Methods and Strategies, Lecture Notes in Computer Science. doi:10.1007/978-3-540-72079-9_8 Surveyed adaptive navigation support as technologies that adapt links to a user's goals and knowledge. It records that these techniques were spreading into adaptive Web applications and adaptive virtual reality.
• De Bra, P., Aerts, A., Berden, B., de Lange, B., Rousseau, B., Santic, T. et al. (2003). AHA! The adaptive hypermedia architecture. Proceedings of the fourteenth ACM conference on Hypertext and hypermedia. doi:10.1145/900065.900068 Presented AHA!, an architecture that began with conditional explanations and link hiding for one online course and grew into a platform whose authoring tools added adaptive features to courses, museum sites and encyclopaedias. Authoring became a first-class design problem here.
• Henze, N. and Nejdl, W. (2002). Knowledge Modelling for Open Adaptive Hypermedia. Adaptive Hypermedia and Adaptive Web-Based Systems, Lecture Notes in Computer Science. doi:10.1007/3-540-47952-x_19 Proposed a knowledge-modelling approach for adaptive open corpus hypermedia, computing ontology subgraphs from standard learning-object metadata to estimate a user's knowledge. Adapting to material the system did not itself author was an explicit design goal.
• Hook, K., Karlgren, J., Waern, A., Dahlback, N., Jansson, C. G., Karlgren, K. et al. (1996). A glass box approach to adaptive hypermedia. User Modelling and User-Adapted Interaction. doi:10.1007/BF00143966 Argued that an adaptive interface risks becoming unpredictable, obscure and uncontrollable, and set the requirement that users be able to inspect the adaptivity mechanisms and control their results. The authors built the POP help system on that principle.
• Kaplan, C., Fenwick, J. and Chen, J. (1993). Adaptive hypertext navigation based on user goals and context. User Modelling and User-Adapted Interaction. doi:10.1007/BF01257889 Built HYPERFLEX, which recommended information from a reader's goals and context using cheap updatable associative matrices, and reported that goal-based adaptive sorting reduced the number of nodes a reader visited. The authors state that the studies were informal and aimed to clarify when adaptation is useful.
• Knutov, E., De Bra, P. and Pechenizkiy, M. (2009). AH 12 years later: a comprehensive survey of adaptive hypermedia methods and techniques. New Review of Hypermedia and Multimedia. doi:10.1080/13614560902801608 Surveyed twelve years of adaptive hypermedia and specified requirements for a new reference model. Its diagnosis is that many systems were tightly integrated with one application or used few techniques, so no generic model could be stated.
• Papanikolaou, K. A., Grigoriadou, M., Kornilakis, H. and Magoulas, G. D. (2003). Personalizing the Interaction in a Web-based Educational Hypermedia System: the case of INSPIRE. User Modelling and User-Adapted Interaction. doi:10.1023/A:1024746731130 Built INSPIRE, in which a learner model carrying knowledge level and learning style drove curriculum sequencing, adaptive navigation support and adaptive presentation, with the learner able to intervene in lesson generation. Both halves of the taxonomy were implemented while the learner could disagree.
• Stash, N. V., Cristea, A. I. and De Bra, P. M. (2004). Authoring of learning styles in adaptive hypermedia. Proceedings of the 13th international World Wide Web conference, Alternate track. doi:10.1145/1013367.1013387 Examined what it takes to author learning-style adaptation, noting that authoring applications were scarce and gave no explicit way to create strategies for particular learning styles. Its objective was a tool linking instructional strategies to adaptive hypermedia.
• Weber, G. and Specht, M. (1997). User Modelling and Adaptive Navigation Support in WWW-Based Tutoring Systems. User Modelling. doi:10.1007/978-3-7091-2670-7_30 Introduced ELM-ART II, a WWW-based LISP textbook in which an overlay model combined with an episodic user model drove adaptive navigation by annotating links and selecting the next curriculum step. Its empirical study reports different effects on different types of users.
• Xie, H., Chu, H. C., Hwang, G. J. and Wang, C. C. (2019). Trends and development in technology-enhanced adaptive/personalized learning: a systematic review of journal publications from 2007 to 2017. Computers and Education. doi:10.1016/j.compedu.2019.103599 Reviewed journal articles on technology-enhanced adaptive and personalised learning from 2007 to 2017, coding parameters of adaptation, learning supports, outcomes, subjects, participants and hardware. It found that preferences, achievements, profiles and learning logs were the main parameters, and that most studies supported only traditional computers.
• Yudelson, M. and Brusilovsky, P. (2008). Adaptive Link Annotation in Distributed Hypermedia Systems: the Evaluation of a Service-Based Approach. Adaptive Hypermedia and Adaptive Web-Based Systems, Lecture Notes in Computer Science. doi:10.1007/978-3-540-70987-9_27 Extended link annotation to a service-based approach in which decision-making is separated from the application and encapsulated in a personalisation service. The paper tests feasibility in the real world and reports an extensive performance evaluation of the architecture.