NMKFR: A Robust Framework for Time-Aware Cold-Start Recommendation

In item cold-start scenarios, recommendation models must balance static item attributes (like text descriptions) with sparse, time-varying behavioral feedback. Existing methods often struggle to coordinate these sources because their reliability shifts as an item moves through its lifecycle. This paper asks: how can a model explicitly measure the uncertainty of temporal evidence to adaptively fuse it with static content for more robust ranking?

NMKFR introduces a dual-branch architecture that separates semantic encoding from temporal state tracking. The semantic branch uses a Titans-based encoder to combine bounded-context attention with long-term memory, preserving both local phrase-level cues and reusable semantic patterns. The temporal branch employs a time-aware Kalman filter to estimate latent states, accounting for irregular interaction intervals.

The core innovation is the Uncertainty-Driven Feedback Mechanism (UDFM). It uses the posterior covariance—a measure of uncertainty from the Kalman filter—to calibrate the system. This uncertainty signal is used in two ways: it modulates memory retrieval in the semantic encoder and adjusts the fusion weights between the static and temporal branches in the Adaptive Comparison Fusion Module (ACFM). By comparing the agreement and disagreement between these branches, the model dynamically shifts its reliance toward the more certain source of information at any given time.

Most cold-start models treat static and temporal evidence as fixed inputs. By treating temporal state estimation as a probabilistic process with measurable uncertainty, NMKFR provides a principled way to handle the "evidence-coordination gap." This allows the model to prioritize static text when behavioral signals are noisy or sparse, and shift toward temporal patterns as more reliable interaction data accumulates. The framework demonstrates that incorporating uncertainty signals leads to stronger ranking performance in offline settings, providing a robust strategy for managing the transition from cold-start to warm-start recommendation.