Classification of NFT Collections

This repository contains the source code, Jupyter notebooks, and experimental results used in the research paper "NFT Collection Classification: A Multimodal Approach Integrating Metadata and Textual Embeddings with Supervised Learning," presented at the 35th Brazilian Conference on Intelligent Systems (Bracis 2025 <link: https://bracis.sbc.org.br/2025/ >).

The project explores different text preprocessing and data balancing techniques for classifying NFT collections into 7 categories.

---

The Experiment

The research and experimentation pipeline, contained in the notebooks, follows these steps:

1. Data Preparation

• Textual data (descriptions of the collections) were processed using different techniques (e.g., stopword removal, lemmatization).

• The target classes are 7 categories:

['art', 'gaming', 'memberships', 'music', 'pfps', 'photography', 'virtual-worlds']

• Different data balancing strategies were tested (e.g., SMOTE, oversampling).

2. Embedding Generation

• A language model from the sentence-transformers library (such as paraphrase-multilingual-MiniLM-L12-v2 or similar) was used to convert the descriptions into numerical vectors (embeddings).

3. Training and Evaluation

• Several machine learning classifiers were trained and evaluated (Random Forest, SVM, Decision Tree, and Logistic Regression).

• The experiments were divided into three main scenarios, the results of which are in the CSVs:

1. results_only_emb.csv: Classification using only text embeddings.
2. results_only_extra.csv: Classification using only statistical data.
3. results_emb_extra.csv: Classification using a combination of embeddings and statistical data.

---

Repository Structure

• /code: Contains the Jupyter notebooks with the entire workflow.

• embeddings_code.ipynb: Main notebook with the pipeline for preprocessing, embedding generation, model training, and evaluation.

• Samuel_Embeddings[Exec-Local].ipynb: Notebook version for local execution and testing.

• /data/results: Contains the detailed results of the experiments.

• Results 7 Categories (BRACIS 2025) - results_emb_extra.csv: Metrics (Accuracy, Precision, Recall, F1, AUROC) for models with embeddings and statistical data.

• Results 7 Categories (BRACIS 2025) - results_only_emb.csv: Metrics for models using only embeddings.

• Results 7 Categories (BRACIS 2025) - results_only_extra.csv: Metrics for the models using only statistical data.

---

How to Reproduce the Experiments

1. Clone this repository and enter the folder:

git clone https://github.com/LABPAAD/nfts_classification.git
cd nfts_classification

2. Start the Jupyter server:

jupyter notebook

3. In your browser, open the /code folder. First, run the embeddings_code.ipynb notebook to process the data and generate the embedding files.

4. After the first notebook is complete, run Samuel_Embeddings[Exec-Local].ipynb to load the embeddings, train the classifiers, and generate the final result CSVs.

---

Application

In addition to this research, a web application (Flask) was developed that consumes the trained model.

• Application repository: (https://github.com/markesley/NFT-Classifier)