[NeurIPS 2025] Source code of "STaRFormer: Semi-Supervised Task-Informed Representation Learning via Dynamic Attention-Based Regional Masking for Sequential Data"

Sequential forecasting pipeline leveraging Gated Recurrent Units (GRU) to model long-term dependencies in ad performance data. A deep learning baseline comparing RNN architectures against TCNs for temporal dynamics.