aerdem4/kaggle-quora-dup

Employs a lightweight LSTM architecture with shared weights and order-invariant feature engineering (min/max frequency, commutative operations) to prevent overfitting on question pairs. Combines 25 engineered features—including NLP metrics, rare word statistics, and LSTM embeddings from GloVe vectors—with 10-fold cross-validation and ensemble averaging, followed by test-set calibration to adjust for class imbalance drift. Built in Keras/TensorFlow with GPU support and includes preprocessing normalization and rare-word replacement to improve generalization.

137 stars. No commits in the last 6 months.