🤖 machine learning

Hierarchical Book Organization for Learning-Resource Discovery using Dual-Path Graph Convolutions

This paper proposes HiGeMine, a hierarchical book classification framework that leverages dual-path graph convolutions to separately model authoritative blurbs and subjective reviews for coarse-to-fine semantic reasoning, thereby improving the organization and discovery of digital learning resources.

Original authors: Suraj Kumar, Utsav Kumar Nareti, Soumi Chattopadhyay, Chandranath Adak, Prolay Mallick

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Suraj Kumar, Utsav Kumar Nareti, Soumi Chattopadhyay, Chandranath Adak, Prolay Mallick

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

In the vast digital libraries of our time, books are no longer just physical objects on a shelf; they are streams of data, each carrying a story and a specific place in the world of knowledge. To find a book that truly fits a reader's needs, whether for a classroom lesson or a quiet evening of reading, we rely on categories. These categories, known as genres, act as a map, guiding us from broad territories like "fiction" or "non-fiction" down to specific neighborhoods like "historical mystery" or "marine biology." However, organizing this digital mountain of text is a difficult task. Unlike movies, which can be sorted by their visual style or music by their sound, books must be understood through their words alone. The challenge is compounded by the fact that a single book can belong to multiple categories at once, and the information available about it often comes in two very different forms: the official, polished description written by the publisher, and the messy, emotional, and sometimes contradictory opinions of thousands of readers.

A team of researchers has developed a new way to solve this sorting problem, creating a system that treats these two sources of information not as a single jumbled mess, but as distinct voices that need to be heard separately before they are combined. They call their system HiGeMine. Instead of trying to force the official description and the reader reviews into a single, uniform analysis, the system first listens to the official description to understand the book's intended purpose. It then uses that understanding to clean up the reader reviews, filtering out the noise of pure emotion or irrelevant complaints to keep only the parts of the reviews that actually describe the story or the subject matter. This process ensures that the system is not misled by a reader who simply says "I loved this" without explaining why, or by a review that focuses entirely on the cover design rather than the content.

Once the information is cleaned and separated, the system begins its work of classification in two steps, much like a librarian first deciding if a book is a story or a fact, and then figuring out exactly what kind of story or fact it is. The first step is a broad decision: is this book fiction or non-fiction? The system makes this choice by looking at the official description and the refined reviews independently, allowing each to speak for itself before combining their insights. If the book is identified as fiction, the system then moves to a second, more specialized set of rules to determine if it is a romance, a thriller, or a fantasy. If it is non-fiction, it looks for categories like history, science, or biography. This step-by-step approach allows the system to handle the complexity of books that fit into many categories at once, a common situation that older methods often struggle to manage.

The researchers tested their system against a wide range of existing methods, including those that rely on simple word counts, complex neural networks, and even the most advanced artificial intelligence models available today. They created a new dataset specifically for this test, gathering thousands of books with both their official blurbs and over 90,000 user reviews. The results showed that their method was significantly more accurate than the others. It outperformed standard machine learning models and even beat the largest, most powerful language models that are currently famous for their ability to understand human language in most categories, though some closed-source models achieved slightly higher scores on specific fiction metrics. The key to this success was not just having more data, but having a smarter way to organize it. By separating the authoritative voice of the publisher from the subjective voice of the reader, and by respecting the hierarchy of genres from broad to specific, the system built a much clearer picture of what each book is actually about.

The study also revealed that the way the system weighs these different sources of information is crucial. It found that for the initial broad decision of fiction versus non-fiction, the official description is very helpful, but for the fine details of specific genres, the filtered reviews provide the most valuable clues. The system learned to balance these inputs dynamically, giving more weight to the source that was most complete and reliable for the specific task at hand. This flexibility allowed it to handle cases where a book might have a very short description but hundreds of detailed reviews, or vice versa. The researchers also discovered that the system's ability to understand how different genres relate to one another—knowing, for instance, that "science fiction" and "fantasy" often share certain traits while "cookbooks" and "biographies" do not—further improved its accuracy.

In the end, this work offers a practical blueprint for organizing the digital world of learning resources. It demonstrates that reliable classification does not require the most expensive or powerful computer models, but rather a thoughtful approach to how information is processed. By acknowledging the different roles that official descriptions and reader opinions play, and by cleaning the data before making a decision, the system provides a scalable foundation for finding the right book for the right reader. This is particularly important for digital learning environments, where students and educators need to navigate vast collections of texts quickly and accurately. The success of this approach suggests that the future of organizing knowledge lies not in treating all text as the same, but in understanding the unique nature of each piece of evidence and letting them work together in harmony.

Technical Summary: HiGeMine – Hierarchical Book Organization using Dual-Path Graph Convolutions

1. Problem Statement

The rapid expansion of digital learning environments has created a critical need for scalable organization and discovery of textual resources, particularly books. While genre information serves as essential semantic metadata for indexing and recommendation, existing classification approaches face three primary challenges:

Flat Classification Limitations: Most methods treat genre prediction as a flat multi-class problem, ignoring the inherent hierarchical structure of literary categories (e.g., Fiction/Non-Fiction as coarse levels, followed by fine-grained sub-genres).
Heterogeneous and Noisy Evidence: Book classification relies on two distinct textual sources: publisher-authored blurbs (authoritative, concise, narrative-focused) and crowdsourced reviews (rich in thematic cues but often noisy, subjective, and sentiment-driven). Directly fusing these sources often introduces semantic interference.
Semantic Discrepancy: There is a fundamental gap between the narrative intent of blurbs and the interpretive, often emotional, nature of reviews. Existing models fail to jointly address the noise in reviews, the semantic discrepancy between sources, and the interdependent nature of hierarchical genres.

2. Methodology: HiGeMine Framework

The authors propose HiGeMine, a hierarchical book classification framework that reformulates genre prediction as a coarse-to-fine semantic reasoning problem. The architecture consists of four core components:

A. Blurb-Guided Semantic Refinement

To mitigate noise in crowdsourced reviews, HiGeMine employs a non-parametric filtering mechanism.

Mechanism: Using a BERT-based encoder, both the book blurb and individual reviews are projected into a shared semantic space.
Filtering: Cosine similarity is calculated between the blurb embedding and each review embedding. Only reviews exceeding a dynamic similarity threshold ( $\Psi$ ) are retained.
Outcome: This process preserves semantically consistent reader evidence while suppressing genre-irrelevant or sentiment-dominated noise, creating a "consolidated review" representation.

B. Semantic-Role-Separated Dual-Path Graph Reasoning

Unlike conventional multimodal approaches that fuse text into a single representation, HiGeMine maintains the distinct semantic roles of blurbs and reviews through independent graph propagation branches.

Graph Construction: Two heterogeneous graphs are constructed: a Blurb-Token graph ( $G_B$ $G_{B}$ ) and a Review-Token graph ( $G_P$ $G_{P}$ ).
- Nodes include document nodes (blurbs or consolidated reviews) and token nodes.
- Edge weights are computed using TF-IDF (for document-token importance) and positive Pointwise Mutual Information (PMI) (for token co-occurrence).
Propagation: The framework uses independent graph convolution layers ( $\Phi_g$ ) for each branch to propagate information. This preserves narrative semantics (from blurbs) and interpretive semantics (from reviews) separately.
Integration: The outputs of the two branches are integrated via a learnable weighting factor ( $\lambda$ ), allowing the model to dynamically balance evidence based on content completeness.

C. Hierarchical Inference (Coarse-to-Fine)

The classification follows a two-stage pipeline:

Level-1 (Coarse): A binary classifier distinguishes between Fiction and Non-Fiction.
Level-2 (Fine): Based on the Level-1 output, a domain-specialized multi-label classifier predicts fine-grained genres.
- Genre-Conditioned Token Semantics: Token embeddings are not static; they are derived from the normalized distribution of word occurrences across specific genre categories, enabling the model to capture genre-aware lexical semantics.
- Label Co-Occurrence Graph: A structured graph ( $G_C$ ) models dependencies between fine-grained genre labels using conditional co-occurrence probabilities. This allows the model to capture inter-genre relationships and overlapping semantics.

D. Training Strategy

Encoders: Pre-trained Language Models (LMs) are fine-tuned for specific tasks and then frozen to stabilize graph reasoning.
Loss: Binary Cross-Entropy (BCE) loss is used for Level-1, while Level-2 uses BCE with logits loss. To prevent error propagation, Level-2 training is gated by ground-truth Level-1 labels.

3. Key Contributions

The paper outlines four primary contributions:

HiGeMine Framework: A novel hierarchical organization framework that disentangles authoritative narrative evidence from subjective crowd-sourced interpretations using a semantic-role-separated dual-path graph architecture.
Blurb-Conditioned Refinement: A mechanism that uses blurbs to contextualize and filter reviews, selectively preserving aligned evidence while suppressing noise, thereby improving robustness against heterogeneous signals.
Genre-Conditioned Representation Learning: A method where token embeddings are derived from their distribution across genre categories rather than static pre-trained lexicons, combined with a label co-occurrence graph to capture inter-label dependencies.
New Benchmark: The curation of a new hierarchical multi-label Goodreads benchmark containing paired blurbs and reviews for 57 fine-grained genres (28 Fiction, 29 Non-Fiction).

4. Experimental Results

The authors evaluated HiGeMine against a comprehensive set of baselines, including traditional classifiers, Multi-Label Text Classification (MLTC) methods, Hierarchical Text Classification (HTC) methods, Graph-based approaches, Pre-trained Language Models (LMs), and Large Language Models (LLMs).

Performance: HiGeMine achieved state-of-the-art results across all metrics.
- Level-1: 96.91% F1-score and 97.50% Accuracy.
- Level-2 (Fiction): 81.89% Micro-F1, 77.24% Macro-F1.
- Level-2 (Non-Fiction): 81.86% Micro-F1, 78.57% Macro-F1.
Comparisons:
- vs. LLMs: HiGeMine outperformed both zero-shot and few-shot open-source (e.g., Llama-3, Qwen) and closed-source (e.g., GPT-4.1, Gemini) LLMs. While few-shot prompting improved LLM performance, HiGeMine remained superior, particularly in non-fiction classification.
- vs. Graph/HTC Methods: It outperformed specialized hierarchical and graph-based baselines (e.g., HiAGM, L-SIMCL, HiGATA) by effectively modeling the distinct semantic roles of blurbs and reviews.
Ablation Studies:
- Removing the review filtering mechanism or the hierarchical structure resulted in significant performance drops.
- The dual-path architecture proved superior to simple feature concatenation.
- Encoder Analysis: While BERT was identified as the most stable encoder overall, the study noted that RoBERTa-L achieved the highest Macro-F1 score (82.29%) for Non-Fiction classification, surpassing BERT's 78.57%, though BERT maintained superior stability across other metrics and levels.
Efficiency: While HiGeMine has higher computational cost than simple baselines due to graph propagation, it is significantly more efficient (approx. 13.6x fewer FLOPs) than large LLMs like Mistral-7B, with an inference latency of ~0.066 seconds.

5. Significance and Claims

The paper claims that HiGeMine provides a scalable foundation for organizing and discovering textual resources in digital learning environments. Its significance lies in:

Reliable Organization: By reformulating genre prediction as a structured reasoning problem, it overcomes the limitations of flat classification and noisy data.
Semantic Integrity: The dual-path approach ensures that the authoritative nature of blurbs and the interpretive richness of reviews are modeled without semantic interference.
Practical Applicability: The framework offers a balance between high predictive accuracy and computational efficiency, making it suitable for large-scale deployment in digital libraries and learning platforms.

The authors conclude that the resulting fine-grained semantic metadata enables more effective resource discovery, while acknowledging that future work could explore multimodal evidence and deeper genre hierarchies.