Introduction
Educational inequality remains a significant global challenge, often rooted in the disparate access to high-quality teaching resources between urban centers and remote regions. While recommendation systems offer a potential technological solution to bridge this gap, standard commercial algorithms—such as those powering Netflix or Amazon—are ill-suited for the classroom. These systems prioritize user preference and viral popularity, whereas educators require materials aligned with specific pedagogical objectives and complex knowledge point dependencies. For instance, teaching multiplication is impossible without a prior understanding of addition.
To address these unique requirements, this paper introduces DFRP (Dual-track Feedback Recommendation), a novel framework designed to provide curriculum-aligned, educationally sound resources by integrating Knowledge Graphs (KGs) with a sophisticated feedback mechanism.
Methodology: The DFRP Framework
The DFRP system is built upon three technical pillars designed to move beyond simple "click-based" metrics:
1. Knowledge Graph Construction and Multi-Modal Fusion
The system maps mathematical knowledge points extracted from standard textbooks (including PEP, SJET, and BNUP series) into a structured Knowledge Graph. This allows the system to respect the hierarchical nature of learning. To understand the content of diverse resources—such as videos, PowerPoints, and lesson plans—DFRP employs a Hypergraph-Driven Multi-Modal Fusion approach. It processes text via ChatGLM, images through EfficientNet-B0, and audio using VGGish. These modalities are fused using a Hypergraph Neural Network, which models higher-order relationships between users and resources that traditional graphs often overlook.
2. Dual-Track Feedback Mechanism
The core innovation of DFRP is its dual-track feedback system, which treats the teacher as an expert evaluator rather than a passive consumer:
- Passive Track: This track monitors traditional engagement metrics, such as click counts and total viewing duration.
- Active Track: This track prompts teachers to complete a "teaching scale" based on Bloom’s Educational Objectives Taxonomy. This scale quantifies a resource’s effectiveness across dimensions like perception, interaction, and cognitive depth, ranging from basic knowledge recall to complex evaluation.
3. Reinforcement Learning Reordering
To ensure the integrity of recommendations, the system utilizes a top-k off-policy correction method. This reinforcement learning component re-ranks final results to mitigate the impact of "malicious clicks" or random noise, ensuring that the most pedagogically relevant materials rise to the top.
Implementation and Key Results
As a demonstration of the system's practical utility, the authors highlight its dynamic adaptation capabilities. In a test scenario where a teacher searches for "angles," the system initially presents a video with a high relevance score (e.g., 0.7047). However, if the teacher provides negative active feedback through the Bloom-based scale after viewing, the system immediately recalibrates.
In subsequent searches, the relevance of that specific video drops significantly (e.g., to 0.5002), demonstrating the system's ability to prioritize pedagogical quality over raw popularity. Furthermore, the platform integrates a visual Knowledge Graph on resource detail pages, allowing educators to see exactly how a specific material links to broader curriculum goals.
Discussion and Future Directions
DFRP represents a shift toward "pedagogy over popularity." By integrating expert evaluation into the recommendation loop, the system ensures that resources are not just engaging, but educationally effective.
However, several areas for future development remain. Currently, the system's Knowledge Graph is focused on mathematics; expanding this to subjects with less linear dependencies, such as literature, will be a key test of the hypergraph model's flexibility. Additionally, while active feedback provides high-quality data, it increases "user friction." Future iterations will need to find ways to streamline this process—perhaps through semi-automated evaluation—without sacrificing pedagogical depth. Finally, large-scale quantitative benchmarking against state-of-the-art models will be necessary to further validate the system's performance across diverse user bases.
Summary updated