PunVas/book-recommender-graphnn-2-tower: v1.0.0

AbstractThis technical report introduces a novel Hybrid Two-Tower Recommendation System designed to address the "cold-start" problem and semantic sparsity in book recommendation. By integrating Graph Signal Processing (GSP) with deep sequential modeling, the architecture effectively fuses global community structure with local semantic features. MethodologyThe proposed architecture introduces a unique pre-processing pipeline:1. Graph-Based Feature Smoothing: Leveraging a low-pass graph filter $(I - \alpha L)^{-1}$ on pre-trained S-BERT embeddings to propagate semantic information across the item-item co-occurrence graph. This is conceptually inspired by the structure-aware updates found in architectures like Evoformer (AlphaFold2).2. Sequential User Encoding: A Gated Recurrent Unit (GRU) coupled with Multi-Head Attention to capture dynamic user interests and temporal evolution.3. Contrastive Optimization: The model is optimized using In-Batch InfoNCE Loss ($\tau=0.07$) to maximize the mutual information between user and item representations in a shared vector space. Experimental ResultsEvaluated on a temporal hold-out set of 26k+ interactions, the system demonstrates superior performance compared to standard baselines:- Recall@10: 86.0%- NDCG@10: 82.9%- Top-1 Accuracy: 79.9% ArtifactsThis repository contains the complete source code, preprocessing scripts, and model architecture definitions. The implementation utilizes PyTorch and SciPy for sparse matrix operations. KeywordsRecommendation Systems, Graph Neural Networks, Graph Signal Processing, Two-Tower Architecture, InfoNCE, Cold Start, S-BERT, Sequential Recommendation.