Evaluation of quantization methods for deep learning modelsProsenc, Urh (Avtor) Demšar, Jure (Mentor) Škrlj, Blaž (Komentor) CTR PredictionModel QuantizationPost-Training QuantizationQuantization-Aware TrainingClick-through rate (CTR) prediction is a cornerstone of modern digital advertising, yet the increasing complexity of models like DeepFM and Deep & Cross Network v2 (DCNv2) introduces significant computational overhead in production environments. This thesis systematically evaluates model quantization as a solution to reduce memory footprint and inference latency while maintaining predictive accuracy. We investigate two primary paradigms: Post-Training Quantization (PTQ) and Quantization- Aware Training (QAT), further augmenting the latter with Knowledge Distillation (KD). Our experiments across three large-scale datasets—Criteo, Avazu, and Teads—demonstrate that while PTQ offers a fast deployment path with minimal accuracy loss in certain configurations (e.g., reaching baseline AUC on the Teads dataset), it is highly sensitive to calibration methods and dataset characteristics. We show that QAT, particularly when resolving framework limitations for custom layers through manual annotation, provides a more robust but computationally intensive alternative. The results provide a comprehensive benchmark for deploying quantized CTR models, highlighting that the optimal quantization strategy is deeply dependent on the specific architectural components and feature distributions of the recommender system.20262026-05-18 10:15:15Magistrsko delo/naloga182581VisID: 38637COBISS_ID: 278916611sl