Higher Education Category Submission

972 – DeepFusionRec – A Hybrid Knowledge Graph & Sequence Behavior Recommendation System

Our system delivers adaptive course recommendations for MOOC platforms by unifying knowledge-graph reasoning with deep sequence modeling. The pipeline comprises five synergistic modules:
Semantic Feature Extraction – A distilled large-language model, DeepSeek-R1-Distill-Qwen-1.5B powered by PaddleNLP, generates compact keyword strings from course metadata. Keywords are embedded into 768-dimensional vectors with ERNIE-3.0-base-zh, accelerated by PaddleNLP + TensorRT FP16 inference for low-latency, high-throughput service.
Text-CNN Behaviour Model – A user’s enrolment history is treated as a “sentence.” Multi-scale 1-D convolutions (kernel sizes 3/4/5, 64 filters each) extract n-gram behavioural patterns from the last ten courses, and a fully connected layer scores every catalogue item. Despite 199 k users and 706 courses, the model trains within 20 epochs and achieves 13.6 % top-1 accuracy on a held-out set.
Knowledge-Graph Construction & Ranking – From 5.6 M triples (ENROLLED_IN, WATCHED_VIDEO, HAS_PREREQUISITE, etc.) we build a heterogeneous MultiDiGraph. Personalized PageRank (α = 0.85) propagates relevance from the target user to courses, ensuring strong coverage for cold-start scenarios.
Dynamic Fusion Strategy – A sigmoid-shaped logistic function adjusts the weights of KG and CNN signals according to the number of courses a user has taken. When interactions are sparse (n ≪ 8), the KG weight ≈ 1; as history grows, emphasis smoothly shifts toward the sequence model while never dropping the KG weight below 0.8 for stability.
Evaluation & Explainability – Offline metrics reach NDCG@10 = 0.503 and Recall@20 = 0.500 for an extreme cold-start user. Integrated visualization tools generate network graphs, path-length histograms, relation-frequency pies, and performance curves, making recommendation rationales transparent to educators and learners.
Impact – The hybrid architecture marries structural domain knowledge with real-time behavioural cues, boosting recommendation quality across the user lifecycle. Its lightweight distillation and FP16 deployment cut inference costs, while explainable graphs foster trust—offering a scalable, transparent blueprint for next-generation educational recommender systems.

Massive Open Online Course (MOOC) platforms have removed geographical and financial barriers, yet they now confront a paradox of choice. Learners often scroll through hundreds of loosely ordered courses, enrol in the wrong level, or abandon study altogether. Interviews reveal two root causes:
1. Cold-start mismatch – New users have little interaction history, so conventional collaborative filtering cannot surface relevant material early enough to spark engagement.
2. Opaque recommendations – Even experienced learners distrust black-box suggestions; without a clear “why,” they hesitate to commit precious time.
Those pain points inspired us to explore a hybrid engine that mixes structural educational knowledge with behavioural patterns, delivering trustworthy guidance from the first click onward.
Our hypothesis is twofold:
• H1 – Coverage & Personalisation: A knowledge graph capturing prerequisite chains, shared core concepts, and peer enrolments can bootstrap relevance for newcomers, while a lightweight Text-CNN can rapidly adapt to emerging behavioural signals. By fusing their scores through a logistic weight that depends on a learner’s activity count, the system should outperform either source alone across the user lifecycle.
• H2 – Trust & Persistence: Visualizing the graph-based paths that connect a learner to each recommendation will increase perceived transparency, thereby reducing the early-dropout rate.
We expect success because (a) our graph already encodes 5.6 M relationships mined from real MOOCs, providing rich semantics unreachable by clicks alone; and (b) it delivers NDCG@10 > 0.50 in extreme cold-start cases.

Steam4

Our solution follows a four-layer workflow that moves raw logs through to a production-ready service:
1. Data & Knowledge-Graph Layer
A Python pipeline parses 1.2 TB of MOOC logs, extracts 5.6 M ENROLLED_IN, HAS_PREREQUISITE, and other triples, and loads them into a NetworkX MultiDiGraph. Status — complete. TSV export available online.
2. Semantic Encoding Layer
For every course, the distilled LLM DeepSeek-R1-Distill-Qwen-1.5B generates condensed keyword strings. These pass to ERNIE-3.0-base-zh, which returns 768-d vectors served via PaddleNLP + TensorRT in FP16.
3. Behaviour Modelling Layer
A lightweight Text-CNN (kernel sizes 3/4/5, 64 filters each) treats a learner’s ten most-recent enrolments as a sentence. Mixed-precision training on 199 k users yields 13.6 % top-1 accuracy and NDCG@10 = 0.503.
4. Fusion & Serving Layer
A logistic scheduler blends CNN logits with Personalized PageRank scores, so the KG weight decays smoothly from ≈ 1.0 (cold start) to a floor of 0.8 (data-rich users). FastAPI + Redis power the REST gateway; Grafana dashboards track latency and quality.
Innovative Features & Implementation Timeline (M0–M6)
• M0–1 — Data audit, KG construction ✔
• M1–2 — LLM distillation, FP16 deployment ✔
• M2–3 — Text-CNN training ✔
• M3–4 — Dynamic fusion engine ✔
• M4–5 — Explainability online dashboard ◀ planned
Performance Metrics
Random personal ranking quality (NDCG@10 ≥ 0.50, Recall@20 ≥ 0.55)

Our recommender fuses cutting-edge deep learning with domain-specific reasoning in a way that is both novel and purpose-built for online education. At its core, we combine PaddlePaddle’s high-performance NLP tool-chain with a hand-crafted knowledge graph of prerequisite links, core themes, and peer interactions. The resulting hybrid engine breaks from the “one-size-fits-all” paradigm that dominates current MOOC platforms.
Paddle-driven semantic distillation. Instead of storing full course syllabi, we ask a distilled LLM, DeepSeek-R1-Distill-Qwen-1.5B, to generate ultra-short keyword strings. PaddleNLP then encodes these strings with ERNIE-3.0-base-zh under TensorRT FP16, shrinking GPU memory requirements to fit on a single A100. This creative compression lets us keep semantic search online for the entire catalogue at negligible cost.
Graph-aware cold-start logic. Traditional recommenders ignore curricular structure; ours elevates it. A 5.6-million-edge knowledge graph is traversed with Personalized PageRank so that, on a learner’s very first visit, suggestions still respect prerequisite chains and shared concepts. The graph signal is not pasted on afterwards; it is blended in real time with Text-CNN logits through a logistic scheduler whose shape adapts to each user’s activity count—an original, mathematically grounded way to transition from exploration to exploitation.
Built-in explainability. We convert the exact PageRank-to-CNN fusion paths into interactive NetworkX-powered visuals, rendered via Paddle.js-compatible WebGL components. Learners can click a recommendation and literally “see” the shortest semantic route from their prior courses to the new one. This creative transparency tackles the trust deficit that plagues black-box AI.
Open innovation loop. Because every element—from ETL to serving—relies on the open-source Paddle ecosystem, educators can fork, audit, or extend the system without vendor lock-in. The synergy of LLM distillation, graph reasoning, and Paddle’s production-grade acceleration therefore represents not just an incremental tweak, but a genuinely creative blueprint for student-centred, trustworthy recommendation at scale.

• Use distilled/quantized models to cut compute and energy.
• Schedule batch jobs in low-carbon periods; scale-to-zero idle services.
Long-term engagement
• Expose recommendation rationales (graph paths, keyword clouds), enable explicit “like/dislike” feedback.
• Introduce progress badges, personalized challenges and milestone rewards.
Adaptive evolution
• Continuous learning pipeline: stream logs → incremental model updates.
• Modular architecture allows hot-swapping new algorithms (e.g. GNNs, causal models) as requirements evolve.
These strategies ensure our system scales with demand, minimizes environmental footprint, fosters enduring user loyalty, and adapts to future needs.

Our recommender tackles a hidden equity gap in online learning: students with limited digital literacy or guidance are far more likely to enrol in courses that mismatch their prior knowledge, leading to early dropout. By injecting prerequisite awareness and transparent explanations into the first recommendation, we give every learner—regardless of background—an informed starting point.
• Prerequisite-aware ranking prevents novices from being funnelled into advanced classes, reducing frustration and widening access to higher-level STEM material.
• Multilingual semantics (ERNIE-zh + English tokenizer) surface Chinese- and English-medium courses side-by-side, supporting bilingual and immigrant learners.
• Explainable graphs demystify why a course appears, empowering users who historically distrust opaque algorithms.

Personal Information Collection Statement (PICS):
1. The personal data collected in this form will be used for activity-organizing, record keeping and reporting only. The collected personal data will be purged within 6 years after the event.
2. Please note that it is obligatory to provide the personal data required.
3. Your personal data collected will be kept by the LTTC and will not be transferred to outside parties.
4. You have the right to request access to and correction of information held by us about you. If you wish to access or correct your personal data, please contact our staff at lttc@eduhk.hk.
5. The University’s Privacy Policy Statement can be access at https://www.eduhk.hk/en/privacy-policy.

• I have read and agree to the competition rules and privacy policy.