TAG-HGT: A Scalable and Cost-Effective Framework for Inductive Cold-Start Academic Recommendation

TL;DR

TAG-HGT tackles the inductive cold-start problem in large-scale academic recommendation by combining semantic recall from an offline LLM-based factory with structure-driven discrimination in a lightweight heterogeneous graph transformer. Using Cross-View Contrastive Learning, the framework distills DeepSeek-V3 semantics into a compact HGT, while a Semantic k-NN graph connects unseen nodes to meaningful semantic neighbors. Evaluated under a Time-Machine Protocol on OpenAlex, TAG-HGT achieves a Recall@10 of $91.97\%$ and attains orders-of-magnitude improvements in inference latency ($4.5 \times 10^{5}$-fold faster) and cost (up to $99.9\%$ cheaper) compared to Generative baselines, through a hybrid scoring $S_{final} = \alpha S_{sem} + (1-\alpha) S_{struct}$ with optimal $\alpha$ near $0.95$. This approach provides a practical blueprint for deploying high-precision, real-time academic recommendations at million-scale, by leveraging Semantics-First, Structure-Refined reasoning.

Abstract

Inductive cold-start recommendation remains the "Achilles' Heel" of industrial academic platforms, where thousands of new scholars join daily without historical interaction records. While recent Generative Graph Models (e.g., HiGPT, OFA) demonstrate promising semantic capabilities, their prohibitive inference latency (often exceeding 13 minutes per 1,000 requests) and massive computational costs render them practically undeployable for real-time, million-scale applications. To bridge this gap between generative quality and industrial scalability, we propose TAG-HGT, a cost-effective neuro-symbolic framework. Adopting a decoupled "Semantics-First, Structure-Refined" paradigm, TAG-HGT utilizes a frozen Large Language Model (DeepSeek-V3) as an offline semantic factory and distills its knowledge into a lightweight Heterogeneous Graph Transformer (HGT) via Cross-View Contrastive Learning (CVCL). We present a key insight: while LLM semantics provide necessary global recall, structural signals offer the critical local discrimination needed to distinguish valid collaborators from semantically similar but socially unreachable strangers in dense embedding spaces. Validated under a strict Time-Machine Protocol on the massive OpenAlex dataset, TAG-HGT achieves a SOTA System Recall@10 of 91.97%, outperforming structure-only baselines by 20.7%. Most significantly, from an industrial perspective, TAG-HGT reduces inference latency by five orders of magnitude ($4.5 \times 10^{5}\times$) compared to generative baselines (from 780s down to 1.73 ms), and slashes inference costs from $\sim$$1.50 to $<$$0.001 per 1k queries. This 99.9% cost reduction democratizes high-precision academic recommendation.

Figures (5)

• Figure 1: The TAG-HGT Framework. (Left) Offline Semantic Factory generates anchors using DeepSeek-V3. (Right) Online HGT Encoder learns structural embeddings. The CVCL module aligns the two views.
• Figure 2: Impact of Residual Weight $\alpha$. The results indicate that a hybrid approach ($\alpha \approx 0.95$) outperforms pure semantics.
• Figure 3: Inference Efficiency.
• Figure 4: Cost Analysis.
• Figure 5: Deployment Architecture. A decoupled serving stack utilizing Redis for low-latency feature lookup, Faiss for approximate nearest neighbor retrieval, and ONNX Runtime for lightweight reranking on CPUs.