## 📖 Competition Description

Modern scientific research spans a wide range of highly specialized domains. Platforms such as **arXiv** organize papers into subject categories to support indexing, discovery, and recommendation. However, manually classifying long and technical research papers is time-consuming and does not scale with the growing volume of submissions.

This competition focuses on **automatic classification of full-length arXiv research papers** into their correct subject categories using machine learning. Unlike short-text classification problems, this task involves **long-form scientific documents**, often containing mathematical notation, domain-specific terminology, and structured arguments.

Participants are expected to design models that go beyond surface-level keyword matching and demonstrate an understanding of **document-level semantics**.

---

## 🔬 Problem Details

Each paper in the dataset belongs to **exactly one arXiv subject category**.

- **Input:** Full text of an arXiv research paper  
- **Output:** One arXiv subject category label  
- **Task Type:** Single-label, multi-class classification  
- **Number of Classes:** 11

Formally, the goal is to learn a function  
Formally, the goal is to learn a function f(text) → category.


that generalizes well across different scientific domains.

---

## 🏷️ arXiv Category Mapping

The dataset uses standard arXiv subject identifiers.  
The mapping from category codes to their full names is provided below:

| Category Code | Full Name |
|--------------|-----------|
| `math.AC` | Commutative Algebra |
| `cs.CV` | Computer Vision and Pattern Recognition |
| `cs.AI` | Artificial Intelligence |
| `cs.SY` | Systems and Control |
| `math.GR` | Group Theory |
| `cs.CE` | Computational Engineering, Finance, and Science |
| `cs.PL` | Programming Languages |
| `cs.IT` | Information Theory |
| `cs.DS` | Data Structures and Algorithms |
| `cs.NE` | Neural and Evolutionary Computing |
| `math.ST` | Statistics Theory |

Each label represents the **primary research area** of the paper.

---

## 📏 Dataset Characteristics

Important properties of the dataset include:

- Papers are **not truncated** and may exceed tens of thousands of characters
- Text may contain:
  - Mathematical expressions
  - Formal definitions
  - Citations and references
- Classes are **not perfectly balanced**

As a result, standard short-text NLP pipelines are often insufficient.

---

## 🧠 Modeling Considerations

Participants are free to use any approach that can run on Kaggle backend. There are no constraints on the approach.

Participants may need to consider approaches such as:
- Transformer-based models with long-context handling
- Chunking or hierarchical representations for long documents
- Aggregation of chunk-level representations
- Classical baselines (e.g., TF-IDF with linear classifiers)

Evaluation is performed using **Macro F1-Score**, encouraging balanced performance across all categories rather than dominance by frequent classes.

---

## 📌 Notes

- Spreadsheet software (e.g., Excel or Google Sheets) cannot reliably open these files due to document length.
- Use Python (`pandas`), R, or similar tools for data loading and analysis.
- The goal is robust document understanding, not heuristic or keyword-based classification.