🤖 AI

RAGEAR: Retrieval-Augmented Graph-Enhanced Academic Recommender

RAGEAR is a neurosymbolic academic recommender system that enhances course suggestions by integrating dense retrieval of fine-grained lecture transcripts with a symbolic Knowledge Graph, utilizing a novel graph-aware aggregation function to outperform metadata-only and transcript-based baselines in ranking quality.

Original authors: Francesco Granata, Lorenzo Lamazzi, Misael Mongiovì, Francesco Poggi, Valeria Secchini

Published 2026-05-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Francesco Granata, Lorenzo Lamazzi, Misael Mongiovì, Francesco Poggi, Valeria Secchini

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a student trying to pick the right university course. You have a specific interest in your head—maybe you want to learn "how to build secure software" or "how to analyze data for legal cases." But when you look at the official course catalog, the titles and descriptions are vague. One course is called "Advanced Programming," another is "Data Structures." Neither title clearly tells you if they actually cover the specific topic you care about.

This is the problem the RAGEAR system tries to solve. Think of RAGEAR as a super-smart, hybrid librarian who doesn't just read the book titles on the shelf; they actually read every single page of every book to find the exact answer to your question.

Here is how it works, broken down into simple parts:

1. The Two Brains: The "Scanner" and the "Map"

RAGEAR uses two different "brains" working together:

The Scanner (The Neural Part): Imagine a robot that has read every single lecture transcript (the text version of the video lectures) from every course. When you ask a question, this robot scans through thousands of pages of text to find the specific sentences or paragraphs that match your idea. It doesn't care about the course title; it cares about the actual words spoken in the classroom.
The Map (The Symbolic Part): This is a giant, structured map of the university. It knows that "Course A" is made of "Lesson 1," "Lesson 2," and "Lesson 3." It also knows the rules: "You need 3 credits," "This is for Computer Science majors," and "You must take Math 101 first." This map ensures the recommendations make sense within the university's structure.

2. The Problem with Just "Summing Up"

In the past, if a robot found 10 matching sentences in a course, it would just say, "This course is good because it has 10 matches."

RAGEAR realized this isn't enough. Imagine two courses:

Course A: Has 10 matching sentences, but they are all crammed into one single 5-minute video clip.
Course B: Has 10 matching sentences, but they are spread out evenly across 10 different lectures throughout the whole semester.

Course B is likely a better fit because the topic is a major part of the whole class, not just a tiny side note. RAGEAR's special trick is a "Graph-Aware Aggregation" function. Think of this as a quality control manager who checks not just how many matches you found, but where they are. It gives a higher score to courses where the relevant information is spread out across many lessons and appears near the top of the search results.

3. How It Works in Practice

You type a question: "I want to learn about the legal side of data privacy."
The Scanner reads: It dives into the transcripts and finds specific sentences in different courses that talk about "GDPR," "data laws," and "privacy regulations."
The Map organizes: It groups these sentences back into their original courses and lessons.
The Score is calculated: RAGEAR calculates a final score. It rewards courses that have:
- A lot of relevant text.
- The most relevant text appearing early in the search results.
- The relevant text appearing in many different lessons (showing it's a core topic, not a fluke).
The Result: You get a list of courses, and the system even shows you the exact sentences from the lectures that made it recommend that course, so you can see why it's a good match.

4. Did It Work?

The researchers tested this system against two other methods:

Method 1 (The Old Way): Only looking at course titles and short descriptions (like reading just the book cover).
Method 2 (The Simple Way): Reading the transcripts but just adding up the matches without checking if they were spread out across the course.

The Results:

Reading the full transcripts (Method 2) was already much better than just reading the titles (Method 1).
But RAGEAR (the full system with the "quality control" scoring) was the best of all. It was particularly good at putting the most relevant courses at the very top of the list.

Summary

RAGEAR is like a personal academic guide that reads the entire library of lecture notes for you. Instead of guessing which course might be right based on a title, it finds the specific moments in the lectures that match your interests, checks if those moments are spread throughout the class, and gives you the best possible recommendation with proof of why it chose them.

Technical Summary: RAGEAR – Retrieval-Augmented Graph-Enhanced Academic Recommender

Problem Statement

Academic course recommendation systems face a significant challenge: students often express learning interests through natural language queries that do not align with the formal terminology found in official course metadata (titles, abstracts, syllabi). Traditional recommenders relying solely on course-level metadata can capture broad thematic similarities but fail to identify fine-grained conceptual evidence present only within actual instructional materials. While online universities possess rich resources like full video lectures and transcripts, exploiting this data requires moving beyond coarse course-level representations to fine-grained retrieval, followed by a mechanism to aggregate this fragmented evidence back into meaningful course-level rankings.

Methodology

RAGEAR (Retrieval-Augmented Graph-Enhanced Academic Recommender) is a hybrid neurosymbolic system designed to address these limitations. It integrates a neural retrieval component with a symbolic Knowledge Graph (KG) to bridge the gap between unstructured lecture content and structured academic constraints.

1. System Architecture

The system operates through four main components:

User Interface: Collects natural language queries regarding academic interests or learning goals.
Knowledge Graph (Symbolic Component): An ontology-based KG models the academic domain, including students, study plans, courses, lessons, credits, prerequisites, and academic disciplines. Crucially, it represents the instructional hierarchy: Course $\rightarrow$ Lesson $\rightarrow$ Transcript Chunk. This structure allows for deterministic symbolic filtering (e.g., by credits or prerequisites) and provides the structural backbone for aggregating retrieval evidence.
Dense Retrieval Module (Neural Component): This module performs dense retrieval over segmented lecture transcripts. Using the multilingual-e5-large sentence-transformer model, it encodes both the student query and transcript chunks (coherent textual units aligned with video timestamps). It returns the top- $k$ (specifically top-200) chunks based on semantic similarity scores.
Recommendation Algorithm: This component propagates chunk-level similarity scores to the course level using a novel graph-aware aggregation function.

2. The RAGEAR Aggregation Function

The core methodological contribution is a scoring function, $RS(C, q)$, which calculates a course-level recommendation score by multiplying three distinct components:

$RS(C, q) = GE(C, q) \cdot RE(C, q) \cdot LC(C, q)$

Global Evidence ($GE$): Measures the share of total retrieved similarity assigned to chunks belonging to a specific course $C$ . This is a normalized SumP aggregation, representing the baseline volume of evidence.
Ranked Evidence ($RE$): Weights chunks based on their rank in the retrieval list. Higher-ranked chunks (those more similar to the query) contribute more significantly. This component uses a smoothing parameter $t_q$ to ensure that highly ranked evidence dominates the score.
Lesson Coverage ($LC$): Evaluates the distribution of relevant evidence across the course's lessons. It rewards courses where relevant chunks are spread across multiple lessons rather than concentrated in a single fragment. This is normalized by the total number of lessons to avoid bias toward longer courses.

This multiplicative approach favors courses that not only have a high volume of relevant content but also possess highly ranked evidence distributed structurally throughout the curriculum.

Key Contributions

Hybrid Neurosymbolic Architecture: The paper presents a system that uniquely combines dense retrieval over full lecture transcripts with an ontology-based Knowledge Graph. This allows the system to leverage fine-grained instructional content while respecting structured academic constraints (e.g., study plans, prerequisites).
Graph-Aware Aggregation Function: The authors propose a novel method for propagating chunk-level retrieval scores to course-level recommendations. Unlike simple summation methods, this function explicitly accounts for the rank strength of retrieved chunks and their distribution across the lesson hierarchy.
Empirical Evaluation: The study provides a rigorous evaluation comparing three approaches:
- Metadata-only dense retrieval.
- Transcript-based normalized SumP aggregation (a baseline using the same retrieval scores but simpler aggregation).
- The full RAGEAR aggregation.

Results

The system was evaluated on 152 student-like queries using a combination of human evaluation (on a subset) and a large-scale LLM-based relevance assessment (using GPT-4.1 nano, validated for high agreement with human raters).

Transcripts vs. Metadata: Utilizing full lecture transcripts significantly outperformed metadata-only retrieval across all metrics (Precision@1, nDCG@5, MAP@5), demonstrating that fine-grained instructional content contains valuable information not captured in course descriptions.
RAGEAR vs. SumP Baseline: The full RAGEAR aggregation further improved ranking quality over the Transcript Normalized SumP baseline.
- Precision@1 increased by 10.69% compared to the metadata baseline and 2.8% over the SumP baseline.
- nDCG@1 improved by 7.21% over the metadata baseline.
- MAP@5 saw a 17.09% improvement over the metadata baseline.
Top-Ranked Performance: The gains were most pronounced for top-ranked recommendations (k=1, 3, 5), indicating that the proposed aggregation function is particularly effective at placing the most relevant courses at the very top of the list.

Significance and Claims

The paper claims that RAGEAR demonstrates the viability of combining symbolic structural modeling with fine-grained semantic retrieval for academic guidance. The authors assert that their primary contribution is not just the use of transcripts, but the specific graph-aware aggregation strategy that effectively synthesizes chunk-level evidence into course-level rankings.

The results suggest that:

Lecture transcripts provide a stronger basis for recommendation than metadata alone.
Aggregating evidence by considering both the rank of retrieved chunks and their distribution across lessons yields superior ranking quality compared to simple summation methods.

The authors maintain a modest stance regarding the evaluation, noting that the LLM-based assessment serves as a controlled proxy for large-scale comparison rather than a substitute for longitudinal user studies with actual students. They emphasize that the symbolic filtering component is deterministic, and the evaluation focused specifically on the ranking component's ability to improve relevance. Future work is directed toward integrating the system into operational services and expanding evaluations to broader academic programs.