How to apply Markov chains for modeling sequential edit patterns in collaborative ontology-engineering projects

Abstract

With the growing popularity of large-scale collaborative ontologyengineering projects, such as the creation of the 11 th revision of the International Classification of Diseases, we need new methods and insights to help project-and community-managers to cope with the constantly growing complexity of such projects. In this paper, we present a novel application of Markov chains to model sequential usage patterns that can be found in the change-logs of collaborative ontology-engineering projects. We provide a detailed presentation of the analysis process, describing all the required steps that are necessary to apply and determine the best fitting Markov chain model. Amongst others, the model and results allow us to identify structural properties and regularities as well as predict future actions based on usage sequences. We are specifically interested in determining the appropriate Markov chain orders which postulate on how many previous actions future ones depend on. To demonstrate the practical usefulness of the extracted Markov chains we conduct sequential pattern analyses on a large-scale collaborative ontology-engineering dataset, the International Classification of Diseases in its 11 th revision. To further expand on the usefulness of the presented analysis, we show that the collected sequential patterns provide potentially actionable information for user-interface designers, ontology-engineering tool developers and project-managers to monitor, coordinate and dynamically adapt to the natural development processes that occur when collaboratively engineering an ontology. We hope that presented work will spur a new line of ontology-development tools, evaluation-techniques and new insights, further taking the interactive nature of the collaborative ontology-engineering process into consideration.

Analyzing user interactions with biomedical ontologies: A visual perspective

“…For example, depending on the category of user [5], as well as his previous browsing behavior, repository developers could provide suggestions on other relevant ontology classes using collaborative filtering algorithms. Similarly, using recurrent neural networks or markov chains [19], repository developers could use the interaction datasets to predict the consequent ontology class the user is likely going to navigate. These predictions and suggestions could then be provided through interactive navigation links.…”

“…In 2013, Wang et al [18] applied association-rule mining to the change-logs of several different collaborative ontology-engineering projects to extract edit patterns, which were then used to predict the next change actions in the corresponding projects. Similarly, Walk et al [19, 20, 21] used (higher-order) Markov chains to study user-editing trails of ontology-engineering projects to predict the action a user is most likely to conduct next. Pesquita et al [22] leveraged the location and specific structural features of edit trails to show that these features can be used to determine where the next change is going to take place in the Gene Ontology.…”

Analyzing User Interactions with Biomedical Ontologies: A Visual Perspective

“…In 2013, Wang et al [18] applied association-rule mining to the change-logs of several different collaborative ontology-engineering projects to extract edit patterns, which were then used to predict the next change actions in the corresponding projects. Similarly, Walk et al [19,20,21] used (higher-order) Markov chains to study user-editing trails of ontologyengineering projects to predict the action a user is most likely to conduct next. Pesquita et al [22] leveraged the location and specific structural features of edit trails to show that these features can be used to determine where the next change is going to take place in the Gene Ontology.…”

How Users Explore Ontologies on the Web: A Study of NCBO's BioPortal Usage Logs

“…Further, there has been increasing interest on analyzing and categorizing human sequences in the past, including mobility [5], page access [7,12], edit [28,30,31] as well as click stream logs [24]. Additionally, Markov models of various orders have been employed [2,15,20,4,22,29] to model and predict clicks or interactions of users on the (Semantic) Web.…”