Online coding education platforms generate large volumes of sequential learner interaction data. Each learner progresses at a different pace, encounters different difficulties, and exhibits unique learning patterns. Traditional recommendation systems often fail to model the temporal evolution of student knowledge, resulting in generic or suboptimal learning paths.
This project addresses the problem of:
Modeling student knowledge evolution over time and recommending personalized next learning activities based on historical learner behavior.
The core objectives are:
- To estimate a learner’s latent knowledge state from interaction sequences.
- To detect learning instability and risk patterns early.
- To recommend the most relevant next coding problems tailored to each learner.
The system is evaluated using a large-scale educational interaction dataset consisting of student-problem attempts collected from an online coding learning environment.
Each interaction sequence includes:
- A student identifier
- A problem (item) identifier
- A timestamp
- A binary correctness label (correct / incorrect)
The dataset is inherently:
- Sequential: interactions are time-ordered.
- Sparse: each learner attempts only a subset of available problems.
- Implicit-feedback based: learner knowledge is inferred from behavior rather than explicit ratings.
Due to privacy, ethical considerations, and dataset licensing, the raw data is not included in the public repository.
To capture temporal learning dynamics, the system employs GRU-based Knowledge Tracing (GRU-KT).
Gated Recurrent Units (GRUs) are chosen because they:
- Efficiently model long-term dependencies in sequential data.
- Are robust to vanishing gradient issues.
- Are well-suited for sparse and noisy educational datasets.
At each time step t, the model receives:
- The problem attempted by the learner.
- The correctness outcome (0 or 1).
These inputs are embedded and fed sequentially into the GRU network.
The GRU hidden state represents the learner’s latent knowledge state, dynamically updated after each interaction.
This state captures:
- Mastered concepts.
- Weak or unstable knowledge areas.
- Learning progression trends.
At each step, the model outputs a probability distribution representing the likelihood that the learner would correctly solve each available problem.
These probabilities are later used for recommendation and learning analytics.
Based on the learner’s most recent knowledge state:
- The model predicts correctness probabilities for all candidate problems.
- Problems already attempted by the learner are excluded.
- The Top-N problems with the highest predicted relevance are selected.
Raw prediction probabilities may be numerically small and difficult to interpret.
To improve usability and explainability, probabilities are normalized within the Top-N set:
[ p_i^{norm} = \frac{p_i}{\sum_{j=1}^{N} p_j} ]
This ensures:
- Relative relevance is clear to human users.
- Percentages sum to 100% within recommendations.
- Ranking quality remains unaffected.
This design choice improves user experience and interpretability without altering the model’s predictive behavior.
To provide transparent insights into learner progress, two complementary metrics are computed.
EMA reflects short-term learning stability, giving higher weight to recent interactions.
- Sensitive to recent mistakes.
- Useful for detecting sudden performance drops.
- Acts as an estimate of current mastery.
Cumulative accuracy represents long-term learning performance, calculated as the running mean of correctness.
- Smooth and stable.
- Reflects overall learning trajectory.
- Less sensitive to short-term fluctuations.
Sharp drops in EMA are identified when:
- The difference between consecutive EMA values exceeds a predefined negative threshold.
These drops often indicate:
- Conceptual misunderstanding.
- Increased difficulty.
- Cognitive overload.
Such points are flagged as risk indicators.
To visualize learner mastery across problems, a Knowledge State Heatmap is constructed.
- For each problem, the average correctness across attempts is computed.
- Problems are ranked by mastery level.
- Color encoding is applied:
- Green → High mastery
- Yellow → Partial mastery
- Red → Weak or unstable knowledge
The heatmap provides:
- An interpretable snapshot of learner strengths and weaknesses.
- Clear identification of problem clusters requiring reinforcement.
- A foundation for targeted intervention and learning plan generation.
The system combines:
- Weak mastery problems (from the heatmap).
- Sudden learning drop points (from EMA analysis).
to generate a personalized learning focus list, where:
- Problems causing both low mastery and sharp drops are marked as high priority.
- Others are marked as medium priority.
This bridges predictive modeling with actionable educational guidance.
The system is intentionally designed as a research-grade prototype, emphasizing:
- Interpretability over black-box predictions.
- Educational relevance over pure accuracy.
- Transparency for instructors, learners, and researchers.
Rather than replacing educators, the system aims to augment decision-making in personalized learning environments.
This methodology integrates:
- Deep sequential modeling via GRU-based Knowledge Tracing.
- Personalized Top-N recommendation.
- Interpretable learning analytics.
- Visual knowledge diagnostics.
By combining prediction and explanation, the proposed system bridges the gap between machine learning models and practical educational insights, making it suitable for academic research, experimentation, and future extension.