UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence

Abstract: Modeling ultra-long user sequences involves a difficult trade-off between efficiency and effectiveness. While current paradigms rely on either item-specific search or item-agnostic compression, we propose UxSID, a framework exploring a third path: semantic-group shared interest memory. By utilizing Semantic IDs (SIDs) and a dual-level attention strategy, UxSID captures target-aware preferences without the heavy cost of item-specific models. This end-to-end architecture balances computational parsimony with semantic awareness, achieving state-of-the-art performance and a 0.337% revenue lift in large-scale advertising A/B test.

• The paper introduces a novel semantic-aware compression using SIDs to bridge the gap between high computation search methods and low computation compression techniques.
• It employs a dual-stage attention mechanism combining explicit semantic probing and gated latent probing to capture both immediate intent and long-term dependencies.
• Empirical results on public and industrial benchmarks demonstrate that UxSID outperforms state-of-the-art methods with a notable 0.337% ad revenue increase and constant-time inference.

UxSID: Semantic-Aware Ultra-Long Sequence Modeling for Recommender Systems

Technical Context and Problem Statement

Scaling personalized recommender systems to ultra-long user behavior sequences is essential for modeling evolving user interests in environments with massive item and user volume, such as short-video platforms and e-commerce applications. Legacy paradigms, including attention-based architectures and memory networks, encounter a computational bottleneck—specifically quadratic complexity—which restricts the deployable sequence length and impedes optimal preference modeling. There exists a dichotomy: search-based approaches (item-specific, high computation) and compression-based methods (item-agnostic, low computation but insufficient for target-aware modeling). This paper introduces UxSID, targeting the unaddressed intermediate regime—semantically-aware compression using semantic identifiers (SIDs), facilitating efficient, precise, and scalable ultra-long sequence modeling (ULSM).

Methodology

SID Generation and Semantic Quantization

UxSID leverages SIDs, discrete codes generated via hierarchical quantization of multimodal content embeddings (derived with a Multimodal LLM and quantized by Res-KmeansFSQ). SIDs encapsulate latent semantic attributes, clustering semantically similar items for shared interest memory routing. By focusing on the first-layer code, UxSID balances semantic granularity and inference latency, supporting scalable production deployment.

Item-Agnostic Interest Compression (IAIC)

IAIC condenses heterogeneous user histories into a set of learnable interest anchors via cross-attention. The anchors are regularized with an orthogonality constraint, maximizing diversity to disambiguate multifaceted user preferences. Per-token feed-forward networks further refine anchor independence.

Hierarchical Semantic Probing

The core innovation is hierarchical dual-stage attention using target SIDs:

• Explicit Semantic Probing: Attends directly to raw behavior sequence using target SIDs, extracting global correlation signals between the candidate item and expansive history.
• Gated Latent Probing: Refines the semantic probing via a gated attention mechanism, conditioning the local probe on global context. The result is a robust fusion capturing both immediate intent and persistent, long-term dependencies.

The concatenated global and local representations yield a target-aware embedding, integrated with user, item, context, and short-term features for final ranking via an MLP head.

Industrial Deployment Pipeline

UxSID is designed for industrial-scale feasibility. The compressed embeddings are precomputed offline and stored via (UID, SID) keys in an embedding server. Real-time serving occurs via $O(1)$ point lookup, ensuring constant latency regardless of original sequence length ($L$ up to 10k+). The system maintains manageable storage overhead due to strong SID clustering properties.

Empirical Evaluation

Performance on Benchmarks

On public datasets (XLong, KuaiRec-Big) and large-scale Kuaishou datasets, UxSID achieves consistent SOTA performance. Notably, UxSID delivers a 0.337% ad revenue increase in production, outperforming robust baselines (SIM, TWIN, C-Former) by significant margins. Importantly, it maintains high robustness against initialization noise and delivers stable gains across sequence lengths.

Component Analysis

Ablations reveal the criticality of SID-driven semantic querying: replacing SIDs with category tags degrades performance and semantic recall. Both the global (eglobal) and local (elocal) stages are indispensable—omitting either leads to substantial AUC drops. Orthogonality loss and gating further improve anchor diversity and query robustness.

Scalability and Efficiency

UxSID achieves constant-time inference per user-item pair, maintaining the highest AUC as the sequence length scales to 10k. Search and static compression baselines display diminishing returns or unstable performance as sequences grow. Hyperparameter sensitivity studies indicate optimal IAIC anchor count (K=16) and moderate orthogonality regularization, avoiding interest entanglement and redundancy.

Visualization and Case Analysis

Attention routing analysis demonstrates SID-driven probes activate relevant interests across the entire behavioral trajectory, outperforming category-based or static compression methods. This activation is closely aligned with click-through objectives, supporting precise intent modeling.

Practical and Theoretical Implications

Practical Impact: UxSID bridges the gap between offline efficiency and online precision, supporting real-time large-scale deployment under stringent latency constraints. Its architecture is well-suited for lifelong behavior modeling, fine-grained interest discovery, and industrial-scale content personalization.

Theoretical Significance: SID-based semantic routing operationalizes the idea of probe-based compression, enabling adaptive, target-aware representations that reconcile global history density with specific intent peaks. The hierarchical attention mechanism combined with anchor diversity regularization sets a new standard for scalable ULSM.

Future Directions: Next steps involve extension to lifelong modeling regimes, leveraging SIDs as universal semantic probes for generative recommendation and further automation of SID codebook generation via LLMs.

Conclusion

UxSID establishes a principled framework for semantic-aware ultra-long sequence modeling by harnessing SID-guided hierarchical attention and efficient item-agnostic compression. The approach outperforms state-of-the-art baselines both in offline and online settings, achieving superior predictive accuracy with industrial-grade efficiency. Its architectural innovations—SID-driven semantic probing and anchor diversity—make it particularly suitable for complex user-item interaction landscapes and future generative recommender paradigms (2605.09040).