We report on an ongoing learning analytics project at the University of Liège, in which we want to analyze student interactions on Cytomine for a histology course. Cytomine provides tools for medical image annotation and an API that has been used for learning analytics. The problem, however, is that the data obtained from Cytomine has implicit semantics and requires many data preprocessing and integration steps. This poster presents the prototype KG we have built to address these problems. The KG adopts PROV-O to distinguish activities from their outcomes, addressing some of the issues faced in the past. We also demonstrate that the KG can be used in Jupyter notebooks, though learning analytics is left for future work. It did demonstrate that the data analysis process has become more declarative and transparent, as data is analyzed starting from SPARQL queries. We focused on one project in Cytomine, and future work consists of integrating additional projects. We also plan to investigate the development of more self-contained KG generation techniques as we have no direct access to the Cytomine application.
CEUR ceur-ws.org
Cytomine [
While advantageous for object persistence, MongoDB’s document-oriented storage model
presents challenges for the interconnected analysis required in learning analytics research.
Additionally, the various document types contain implicit relationships, so one must manually
determine a user’s subsequent annotations, for example. As such, prior learning analytics
studies [
This paper briefly discusses our approach to integrating Cytomine’s data into a KG,
demonstrates our KG in a Jupyter Notebook, and elaborates on future work. The potential of this
study is substantial, as the feedback provided to students will guide their studies and enhance
their performance. Moreover, the data will assist educators in efectively integrating digital
microscopy into their pedagogical plan, thereby optimizing educational outcomes.
1.1. Related Work
There is little related work on the use of KGs for learning analytics. The learning analytics
community seems to focus on using Linked Data to facilitate research, as can be observed in
the LAK Data Challenge [
Ontology Development The KG’s ontology was engineered with a middle-out approach
where entities in the data (described below) were identified and aligned with the UoD of
domain experts and existing ontologies. We adopted OWL 2 QL as we anticipate the KG to
contain many assertions. The ontology we developed builds upon PROV-O [
Data Transformation We had no access to Cytomine’s MongoDB instance, though we could
download the data via its API.2 The data of one project consisting of 11 images, 588 users
(pseudonymized), and 27185 annotations, 1571 properties, and 31507 descriptions. We
used RML [
As such, we enriched the data with a geometric dimension.
We recognize that our approach’s major limitation is its inability to transform the data stored in MongoDB. Moreover, Cytomine’s API is fairly restricted, allowing us to retrieve data when suficient restrictions are placed (e.g., retrieving the annotations on a project-per-project basis). This limitation is beyond our control.
The result of this study yielded a proof-of-concept KG for learning analytics. The KG can be
explored with tools such as Ontodia [
To demonstrate that one could engage with the KG for learning analytics, we created a Jupyter Notebook that retrieved the number of annotations per contributor and used this to determine the optimal number of clusters using the Elbow Method, as shown in Figure 3.
We reported on the feasibility of creating a KG out of Cytomine, which required integrating CSV into RDF. The data we obtained from Cytomine was rather flat. Information about a user’s 1The ontology, available at https://chrdebru.github.io/papers/2024-09-semantics/ontology.owl, is not yet made available using a persistent identifier. The ontology will be published in a future iteration of the KG construction. 2https://doc.uliege.cytomine.org/dev-guide/api/reference activity was implicitly stored but rendered explicit using PROV-O in the KG generation process. As users annotated slides and stored them with geometric coordinates, we adopted GeoSPARQL to use geospatial predicates. This allows us to analyze interactions on specific regions on slides, for example. The number of annotations within one project indicates our project’s scale, knowing there are over 150 projects in Cytomine. Challenges that we will investigate include the evolution of this KG over time. As we currently have no access to the MongoDB instance, which is normal, we should investigate more elegant ways to generate the KG. One venue is to retrieve the data via rest calls in the mapping, which requires the development of bespoke RML iterators.