Overview
LMS platforms record every meaningful learning touchpoint—from viewing resources and submitting work to discussions and instructor feedback. My research turns those signals into actionable insights for online and blended course design. Using the Community of Inquiry (CoI) framework (teaching, social, cognitive presence) and Moore’s learner–content, learner–learner, and learner–instructor interactions, I model how different engagement patterns relate to student satisfaction and to achieving a competency threshold of 70%.

Across fully asynchronous graduate courses, I identify interaction archetypes (e.g., content-heavy, discussion-oriented, balanced) and show how course type matters: knowledge-based courses benefit from stable content engagement, while application-based courses thrive when instructor feedback cycles and peer interaction are stronger. I also rank predictors of success (such as timely submissions, forum participation cadence, and feedback touchpoints) to inform interventions.

These findings inform data-driven personalization, from assignment pacing and forum structure to analytics dashboards that faculty can effectively utilize, enabling programs to boost engagement, increase satisfaction, and enhance performance at scale.

Grounded in Theory

Community of Inquiry (CoI) diagram
My work uses the Community of Inquiry framework to show how learning emerges where Teaching presence, Social presence, and Cognitive presence meet. The overlaps highlight Setting Climate, Supporting Discourse, and Regulating Learning, with the Educational Experience at the center. In fully asynchronous graduate courses I translate these presences into LMS signals such as instructor feedback cadence, discussion participation, and resource engagement, and I link them to student satisfaction and achieving a ≥70% competency benchmark.

Framework combining Dewey pragmatism, CoI, and Moore interactions
This layered model connects learning philosophy to measurable practice. The outer ring reflects Dewey’s pragmatism, which views learning as a purposeful inquiry and an experiential process. The next layer applies the CoI presence to define the conditions that support that experience. The inner layer utilizes Moore’s interactions: learner-to-content, learner-to-learner, and learner-to-instructor to provide the operational channels captured in the LMS. Together, the model explains how specific interaction patterns stimulate meaningful inquiry, align with course type in knowledge-based and application-based courses, and predict outcomes that guide data-informed course design.

Methodology

How the analysis was done
I transformed LMS event logs into learner level features that represent real activity and presence, including content views per week, time on task, on time submissions, discussion posts and replies, announcement opens, and instructor feedback touches. Features were cleaned, winsorized where needed, and standardized so that one metric could not dominate another. I then used a sequence of analyses to discover patterns, relate them to Curriculum and Instructor Evaluation satisfaction, and predict competency at or above seventy percent.

Results

Dataset and preprocessing
Nine fully asynchronous graduate courses (labeled A–I) were analyzed. LMS activity features (content views per week, time on task, on-time submissions, discussion posts and replies, announcement opens, instructor feedback touches, quiz attempts, etc.) were z-score standardized prior to analysis.

Principal Component Analysis
Strong inter-variable correlations motivated the use of PCA, as evidenced by correlations between Content Object Count and Quiz Count (r = 0.92) and Dropbox Submissions (r = 0.95). The first components summarized content engagement, discussion/peer exchange, and instructor engagement, and the component scores were used for clustering. See PCA score plot and PCA biplot figures.

K-Means Clustering

K-means clustering on standardized LMS features and PCA scores yielded a three-cluster solution that strikes a balance between fit and interpretability.
Cluster 0 mixed interaction: E, G, H. A heterogeneous profile across metrics with some above median signals and others below median, indicating course-specific choices rather than a single pattern.
Cluster 1 assignment focused on high engagement: F. Strong assessment activity with steady participation signals and ample content delivery.
Cluster 2 assessment intensive and instructor-driven: A, B, C, D, I. High content access, frequent quizzes and grading activity, and comparatively low discussion.

Content related metrics such as content objects, quiz activity, and grade producing tasks are highest in Cluster 2, indicating consistent access to materials and assessments. Cluster 1 also shows above median content delivery alongside its assignment cadence. Cluster 0 shows the most variation within the group, with one course above median on several content measures and the other two below median, which points to course level design differences rather than a shared cluster wide pattern.

Peer to peer engagement is limited in Clusters 1 and 2, where discussion activity remains low even when overall engagement is high. Cluster 0 shows the widest spread on discussion forums and topics. One course in this cluster contributes most of the peer exchange, while the other two remain near zero. This indicates that opportunities for sustained dialogue are uneven and largely determined by course design choices rather than by overall activity level.

Instructor activity is strongest in Cluster 2, which combines frequent grading activity and feedback touches with structured assessment cycles. Cluster 1 also shows above median instructor involvement aligned with its assignment rhythm. Cluster 0 presents lower and more uneven instructor interaction signals, again suggesting course specific rather than cluster wide emphasis.

Application-Based Course vs Knowledge-Based Course

Application-based offerings are concentrated in the assessment-intensive, instructor-driven profiles, which comprise Cluster 2 (A, B, C, D, I), as well as the single-assignment-focused, high-engagement course in Cluster 1 (F). These courses show elevated content access, frequent quiz/assignment cycles, and a tighter grading/feedback cadence.

Knowledge-based offerings appear in the mixed-interaction profile (Cluster 0: E, G, H) and in one assessment-intensive case among A–D, I, indicating greater heterogeneity. In Cluster 0, knowledge-based courses emphasize steady content viewing with fewer assessment events and more variable instructor touches.

Across both types, discussion volume is generally low. The largest peer-exchange signal occurs in Cluster 0, driven primarily by a single course, while other offerings, regardless of type, remain near minimal discussion levels.

Utilization vs Course Satisfaction

At the course level, CIE satisfaction is not positively associated with simple volume metrics. Several activity counts show negative correlations with satisfaction: Dropbox submissions r = −0.84, grade count r = −0.65, and content object count r = −0.49. Discussion and assessment volumes are weakly negative: discussion forums r = −0.24, quiz attempts r = −0.22. Discussion topic count is approximately zero (r = 0.043). Overall, higher activity volume does not correspond to higher satisfaction; satisfaction appears to depend on the structure and timing of activity rather than raw counts.

LMS Utilization vs Performance

Benchmark status
All courses met the program benchmark of 70 percent competency. Courses A, B, C, E, and G reached 100 percent. Courses D, F, H, and I were between 92.9 percent and 99 percent.

Model results
A Random Forest model identified the strongest utilization signals associated with meeting the benchmark. The top features were total time in content, percentage of content completed, discussion posts read, total quiz attempts, and number of logins.

Cluster comparison
Competency rates were uniformly high across clusters. Differences by cluster were small because every course cleared the benchmark.

Top Predictors of Performance

A Random Forest classifier with stratified cross-validation identified the strongest signals linked to meeting the 70 percent competency benchmark.

Top features

1. Total time in content
2. Percentage of content completed
3. Discussion posts read
4. Total quiz attempts
5. Number of logins

Recent Articles

Article in Online Learning Consortium

Article in eLearning Industry

Future Research: Applying LSTM Deep Learning to LMS Interaction Data

As part of my ongoing work in learning analytics and the Community of Inquiry (CoI) framework, my next research direction focuses on leveraging Long Short-Term Memory (LSTM) neural networks to analyze temporal sequences of learner interaction data within Learning Management Systems (LMS).

While my current studies employ clustering, PCA, and Random Forest classification to uncover engagement patterns, LSTM models offer a deeper way to capture time-dependent learning behaviors—revealing how engagement evolves across course weeks, modules, or key assessments.

Through this approach, I aim to:

• Detect early indicators of learner disengagement or achievement.

• Map behavioral trajectories linked to social, cognitive, and teaching presence.

• Integrate LMS logs with affective data, such as discussion sentiment or feedback tone.

• Develop adaptive feedback mechanisms that guide instructors in real time.

This line of research bridges deep learning and educational theory, moving toward predictive, personalized insights that support meaningful online learning experiences.