Core logic: Deciding what to study next

Jon Fernquest, Mae Fah Luang University, IT Department, 5/11/2021

[Full Index]

Study recommendation, addressed in this study, is new and different. 

Knowledge-based recommendation systems (also called ‘Ontology-based’)  typically recommend purchases to online shoppers, friends on social media or content-recommendation such as movies, books or academic articles (Aggarwal 2016:167-197). They employ one of two approaches, constraints or cases, of which only the case-based approach is relevant here.

Study recommendation recommends items for further study in a given learning domain at a given point during a course of study. Content recommendation is the closest of the above-mentioned common systems to this type of system. The problem of study recommendation, however, is directed to a much more narrowly defined domain than content recommendation, focusing on specific facts and items of information in this domain, and is thus inherently quite different. Drawing analogies with the more common purchase, friend and content recommendation is an essential first step in defining how study recommendation functions.

The basic problem that ‘knowledge-based’ recommendation tackles is helping a user navigate through a complex and unfamiliar knowledge domain. The aim is to make some complex choice or decision in an optimal fashion, such as the purchase of real estate in a given locale or finding an academic article relevant to one’s research.

In the study recommendation case, the problem is finding the next item to study and review in a textbook or syllabus. This is done with the goal of achieving a high test score or high assessment in a course or towards the broader goal of mastering that particular knowledge domain. The base case proceeds as follows. In a course, learning material is presented for the first time by the instructor. The study recommendation system then helps the student to study and review this material. Finally, some form of assessment takes place for which the student receives a grade.

After the initial presentation of learning material occurs, spaced repetition at intervals of increasing length for speedy retrieval of facts from memory is essential so the student does not forget the material. However, in answering the question of what to review next, spaced repetition isn't the only factor and is arguably not the most important factor. The most important factor would be to what degree the student has mastered or grasped different segments of the syllabus and learning material and what remains for the student to master and grasp. 

What the student has mastered is not measured directly by the student rating items as it is in purchase and content recommendation systems, but rather measured indirectly by two instruments: responses to flashcards and test questions (multiple choice, matching, true-or-false, short-answer or drag-and-drop).  Flashcards will be used in the discussion that follows but the points apply equally to the test-question instrument as well.  In test questions, the score received indicates the need for further review.

Flashcards & Self-Assessment

The essence of the flash card is self-assessment. The flashcard has two sides: front and back.  The front side may be a question with the back side an answer, or front and back may correspond to ideas, definitions or concepts for the student to master.  When the student is presented with the front side, the student should be able to map it to the back side in their mind. The student sees the front side, tries to recall the backside from memory and then flips to the back side to check and self-assess their competency or mastery on a scale from 1 to 5 (as is the case with the Brainscape flashcard app).

Through the self-assessment of flashcards, the potential utility of further study of the fact and knowledge on this particular flashcard and other related flashcards is revealed to the study recommendation system. 

Knowledge Graphs

Another essential part of the study recommendation system is the knowledge-base (KB) of learning material which comes from the syllabus and textbook of the course. Given a particular fact or item of knowledge, the KB has the potential to reveal further related facts.

Since Google popularized them in the early 2010s, so-called ‘knowledge graphs’ have become the de facto standard for representing knowledge bases. In the case of dental anatomy that we address here, these knowledge graphs incorporate very specific relations between specific named-entities in the dental anatomy knowledge domain. This includes the specific features (or ‘named entities’) of teeth such as the cusps, ridges, grooves, pits or curvatures of, let’s say, a molar.  The location and spatial relations between these features are described using standard medical terminology such as mesial, distal, lingual, buccal, facial, cervical third of buccal surface of crown…etc. These verbalized descriptions are most importantly associated with visual features to be recognized. This means that almost any flashcard or test question can fruitfully augmented with an image to aid student mastery (unless the objective is to have the visualize in their own mind without the aid of images).

So, to summarize, we have two essential components for the study recommendation system: 1. knowledge graphs of dental anatomy knowledge in the dental anatomy knowledge base, and 2. student self assessments of whether they understand and grasp particular items, features and relations in this knowledge base.

Similarity Metrics & Utility for Further Study

The essential part of the case-based approach to recommendation systems is to apply a similarity metric to knowledge graphs in the knowledge-base.

The flashcard just studied is mapped to its knowledge graph which in turn is mapped to related knowledge graphs and their flashcards for recommendation for further study.

Flashcards feature student self-assessments of whether they grasp the fact on the flashcard. These student self-assessments (or hard scores, for test questions) are treated as a utility measure for further study. 

The flashcard self-assessment is mapped to a utility for further study. Self-assessed high-mastery maps to low-utility for further study. Self-assessed low-mastery maps to high-utility for further study. 

This utility for further study is also applied to related knowledge graphs and flashcards to assess their utility for further study as well.

Thus self-assessment applied to the knowledge base produces additional recommendations. For instance, if student is has having problems with identifying curvature, this might lead to recommending the study of specific curvatures on other teeth as well. Studying a feature or relation on one type of tooth leads, via knowledge graphs, to studying the same on other types of teeth. Inter-tooth comparisons are in fact an important part of the dental anatomy knowledge base, corresponding to how dental anatomy is taught in Woelfel’s important textbook and other textbooks on dental anatomy in which, for instance, molars are first compared with other molars and then with other teeth.  

If the student has having problems mastering certain features or relations, a high further study utility metric is calculated, and the knowledge graph is traversed revealing additional related potentially problematic features and relations which are presented as recommendations to study.

Students might prefer to go through flashcards first and then quiz items second or they might just prefer quiz questions and hard scores only, or they might prefer certain types of flashcards and certain types of quiz questions. 

In summary, study recommendation systems can be defined in an analogous fashion to the more common recommendation systems. They follow the patterns of knowledge-based recommendation systems, but they do so in a much more narrowly defined and detailed knowledge domain in a given course of study.

REFERENCES

Aggarwal, Charu C. Recommender Systems: The Textbook. Springer. 2016.