ecoicop-pt-cnn

🧾 ECOICOP Product Classifier (CNN)

This model classifies short product titles (product + brand) into 71 ECOICOP food-and-beverage categories in Portuguese.
It was trained on a manually verified dataset collected from six major Portuguese retailers.

---

📦 Model Overview

---

📊 Test-Set Performance (20 % hold-out)

---

🚀 Quick Start

1 — Install dependencies

pip install torch joblib scikit-learn huggingface_hub

2 — Predict in Python

import torch, joblib, torch.nn as nn
from huggingface_hub import hf_hub_download

# ── Same architecture used during training ──
class SimpleCNN(nn.Module):
    def __init__(self, input_dim, num_classes):
        super().__init__()
        self.fc1 = nn.Linear(input_dim, 512)
        self.relu = nn.ReLU()
        self.dropout = nn.Dropout(0.3)
        self.fc2 = nn.Linear(512, num_classes)
    def forward(self, x):
        x = self.fc1(x)
        x = self.relu(x)
        x = self.dropout(x)
        return self.fc2(x)

repo_id = "julisses/ecoicop-pt-cnn"

# Download vectorizer, label encoder and weights
vectorizer    = joblib.load(hf_hub_download(repo_id, "vectorizer.pkl"))
label_encoder = joblib.load(hf_hub_download(repo_id, "label_encoder.pkl"))
state_dict    = torch.load(hf_hub_download(repo_id, "model.pt"), map_location="cpu")

# Re-create and load the network
model = SimpleCNN(len(vectorizer.get_feature_names_out()), len(label_encoder.classes_))
model.load_state_dict(state_dict)
model.eval()

def predict_ecoicop(text: str) -> str:
    X = vectorizer.transform([text]).toarray()
    with torch.no_grad():
        idx = model(torch.tensor(X, dtype=torch.float32)).argmax(1).item()
    return label_encoder.inverse_transform([idx])[0]

# Example
print(predict_ecoicop("Queijo Mini Original, Cheddar e Emmental Babybel"))
# ➜ Queijo e requeijão

---

✅ Intended Use

• Automatic classification of Portuguese retail product titles (product + brand)
• Supports statistical offices, inflation research and retail analytics

---

⚠️ Limitations

• Trained only on food & beverage products
• Not suitable for unrelated items or non-Portuguese text
• Seasonality gap – the training data cover only a single snapshot in time, not a full calendar year.
  Categories that contain strong seasonal items (e.g. Easter sweets, Christmas pastries, summer beverages) may therefore be under-represented, so prediction accuracy can drop for those products outside the sampled period.

---

📄 License

• Code — MIT
• Weights — © Juliana Machado, released for research & non-commercial use