githubDCS/belgo-classifier

BelgoClassifier - Belgian Transaction Categorizer

CoreML model for categorizing Belgian bank transactions into 15 categories.

Model Details

• Version: 1.0.0
• Accuracy: 83.8%
• Framework: CoreML (iOS/macOS)
• Input: 384-dimensional embedding vector
• Output: 15-class probability distribution

Categories

1. groceries - Supermarkets, food stores
2. restaurants - Restaurants, fast food, delivery
3. transport - Public transport, fuel, parking
4. utilities - Electricity, gas, water
5. telecom - Phone, internet providers
6. healthcare - Medical, pharmacy, insurance
7. insurance - All insurance types
8. housing - Rent, mortgage
9. entertainment - Cinema, events, sports
10. shopping - Retail, online shopping
11. subscriptions - Streaming, software
12. income - Salary, refunds
13. transfers - Bank transfers
14. cash - ATM withdrawals
15. other - Uncategorized

Usage in iOS

The model expects a 384-dimensional embedding vector (from sentence-transformers/all-MiniLM-L6-v2 or Apple's NLEmbedding).

// Load model
let model = try MLModel(contentsOf: modelURL)

// Create input
let embedding = getEmbedding(for: transactionText) // 384-dim vector
let input = try MLDictionaryFeatureProvider(dictionary: ["embeddings": embedding])

// Predict
let output = try model.prediction(from: input)
let probabilities = output.featureValue(for: "probabilities")

Training

Trained on synthetic Belgian transaction data including:

• Belgian supermarkets (Colruyt, Delhaize, Carrefour, etc.)
• Belgian banks and insurers
• Belgian telecom providers
• Common Belgian merchants

License

MIT License - Free to use in commercial applications.

Links

• App: BelgoBudgetto (iOS)
• Training code: Private repository