SGP-Rec: Non-linear Spreading Activation with Degree-Discounted Conductances for Cold-Start Interactive Recommendation
Authors/Creators
Description
We study interactive cold-start recommendation on a bipartite user–item graph, where newly arriving users are presented with slates of (K=6) items and respond through watch time, likes, and superlikes. We identify a structural limitation of degree-normalized one-hop methods, including item-based cosine similarity and RP3β: items bridging distinct communities exhibit inflated degrees, causing normalization to suppress precisely those items required for niche discovery. Through a systematic sweep of the RP3β exponent and a factorial ablation study, we establish an empirical one-hop trade-off between hub suppression (hit rate) and bridge crossing (discovery).
We introduce SGP-Rec, a nonlinear spreading-activation framework combining graded bipolar conductances, incrementally maintained user potentials, degree-discounted impedance, and sparse two-hop diffusion selectively restricted to the Top-B non-massive bridge items. On a bottleneck benchmark, SGP-Rec achieves 100% discovery, with a 97.5% Wilson 95% lower confidence bound, while maintaining a 91.6% hit rate—a performance configuration unattained by any tested one-hop variant. On a continuous latent manifold under Zipf-distributed exposure, SGP-Rec increases long-tail exposure by 22.2% relative to item-CF, while remaining competitive in hit rate. All reported evidence derives from synthetic data generators; the scope and limitations of this experimental setting are explicitly stated.
Files
pirolo2026_sgp_rec_v3.pdf
Additional details
Software
- Repository URL
- https://zenodo.org/records/23898526
- Development Status
- Active