hm-fashion-recommendation

H&M Fashion Recommendation Model

Model Description

A LightGBM binary classification model converted to ONNX format for personalized fashion recommendations. Built for the H&M Personalized Fashion Recommendations Kaggle competition (MAP@12 metric).

The model predicts whether a customer will purchase a specific article based on 30 engineered features derived from customer behavior, article properties, and transaction history.

Model Details

• Architecture: LightGBM (gradient boosted decision trees)
• Format: ONNX (converted via OnnxMLTools)
• Task: Binary classification (purchase / no-purchase)
• Input: 31 features (float32), shape [batch_size, 31]
• Output:
  • label: Predicted class (int64)
  • probabilities: Class probability distribution
• Trees: 200 boosting iterations
• Objective: Binary sigmoid

Input Features

The model expects 31 features including:

• Customer demographics (age, club membership, fashion news)
• Article attributes (price, garment group, product type, color, department)
• Transaction-derived features (purchase counts, recency, frequency)

Usage

import onnxruntime as ort
import numpy as np

# Load model
session = ort.InferenceSession("H&MRecommendationSystemModel.onnx")

# Prepare input (shape: [1, 31])
input_data = np.random.rand(1, 31).astype(np.float32)

# Run inference
inputs = {"input": input_data}
label, probabilities = session.run(None, inputs)

print(f"Predicted label: {label[0]}")
print(f"Probabilities: {probabilities}")

Training

• Framework: LightGBM
• Dataset: H&M transaction data (Kaggle)
• Feature Engineering: 15 new features engineered, 11 dropped from raw data
• Cleaning Strategy: Median version (best of 4 strategies tested)
• No image data used (resource constraint)

Limitations

• Trained on H&M-specific data; may not generalize to other retailers
• Does not use image features (tabular only)
• Binary classification per (customer, article) pair — ranking requires post-processing

License

Apache License 2.0