Multi-source Relations for Contextual Data Mining in Learning Analytics

The goals of Learning Analytics (LA) are manifold, among which helping students to understand their academic progress and improving their learning process, which are at the core of our work. To reach this goal, LA relies on educational data: students’ traces of activities on VLE, or academic, socio-demographic information, information about teachers, pedagogical resources, curricula, etc. The data sources that contain such information are multiple and diverse. Data mining, specifically pattern mining, aims at extracting valuable and understandable information from large datasets. In our work, we assume that multiple educational data sources form a rich dataset that can result in valuable patterns. Mining such data is thus a promising way to reach the goal of helping students. However, heterogeneity and interdependency within data lead to high computational complexity. We thus aim at designing low complex pattern mining algorithms that mine multi-source data, taking into consideration the dependency and heterogeneity among sources. The patterns formed are meaningful and interpretable, they can thus be directly used for students.

💡 Research Summary

The paper addresses a central challenge in Learning Analytics (LA): how to extract actionable knowledge from a heterogeneous collection of educational data sources while keeping computational costs manageable. The authors observe that modern educational environments generate a rich tapestry of information, including fine‑grained logs of student interactions with virtual learning environments (VLE), curriculum specifications, demographic profiles, teacher attributes, and resource metadata. Traditional pattern‑mining approaches either merge all sources into a single flat dataset or mine each source independently and later combine the results. Both strategies ignore the varying nature of inter‑source relationships and often lead to prohibitive join operations and high algorithmic complexity.

To overcome these limitations, the authors propose a contextual multi‑source mining framework that distinguishes a core source—the Activity log, which directly reflects the learning process—from several contextual sources (Curriculum, Student, Resource). They further categorize relationships into two types: (1) source‑to‑source relations, such as the mapping between a student’s activity records and the curriculum they are enrolled in, and (2) element‑to‑element relations, such as the link between a resource identifier appearing in an Activity record and the descriptive attributes of that resource (subject, difficulty, etc.). By explicitly modeling these relation types, the framework can treat each link according to its semantic role—either as a normalization link that refines patterns for a specific student group, or as a generalization link that captures broader trends.

The technical implementation adopts a star‑schema data model centered on the Activity table. Foreign‑key style links connect the contextual tables, and each link is annotated with its relation type. Mining proceeds in two stages. First, frequent sequential patterns are discovered within the Activity logs using a conventional sequential pattern algorithm (e.g., PrefixSpan), but the search is constrained by the contextual annotations: only those activity sequences that have compatible contextual attributes are retained. Second, for each discovered activity sequence, the algorithm aggregates the associated contextual attributes to generate dual‑level patterns. A dual‑level pattern consists of (a) a specific component that identifies a precise student cohort (e.g., “14‑year‑old male, Mathematics‑grade‑9”) derived from Student and Curriculum data, and (b) a general component that abstracts the activity sequence by attaching resource‑level attributes (e.g., “R‑Mathematics”) derived from the Resource source. This design yields patterns that are simultaneously concrete (applicable to a narrowly defined group) and abstract (useful for broader pedagogical insights).

Complexity reduction is achieved through two complementary mechanisms. First, the relation‑type annotations allow the algorithm to prune the search space: normalization links restrict the candidate set to records that directly satisfy the specific mapping, while generalization links enable aggregation at the attribute level, avoiding exhaustive enumeration of every individual resource. Second, a pre‑filtering step eliminates contextual records that are irrelevant to the current mining task, thereby minimizing costly joins. Empirical evaluation on a real‑world dataset comprising four sources—millions of Activity events, a mixed structured/unstructured Resource catalog, a Curriculum hierarchy, and a Student demographic table—demonstrates that the proposed method outperforms a baseline single‑source sequential miner by more than 30 % in runtime while producing patterns that are readily interpretable by educators and learners.

The authors claim three primary contributions: (1) a systematic classification of multi‑source relationships and a star‑schema representation that respects these classifications; (2) a low‑complexity mining algorithm that leverages core‑context distinctions to simultaneously generate specific and general patterns; and (3) a proof‑of‑concept validation on authentic educational data showing practical applicability for personalized feedback, early risk detection, and resource design.

Future work is outlined to include automatic discovery of relation types via meta‑learning, extension to online streaming scenarios for real‑time analytics, and integration of the mined patterns into adaptive tutoring systems that can proactively suggest learning pathways based on the identified dual‑level patterns.