E-commerce search relevance modeling faces a critical dilemma: traditional models falter with complex queries, while Large Language Models (LLMs), despite their superior reasoning, suffer from the prohibitive latency of auto-regressive Chain-of-Thought (CoT) generation, rendering them infeasible for production. Knowledge distillation offers a promising solution, yet current methods force an undesirable trade-off: sacrificing the very interpretability that makes LLMs powerful, or relying on expensive, unscalable human-annotated rationales. To address this, we propose STAR—Staged Training with Aligned Reinforcement Learning and Multi-Faceted Distillation, a progressive framework that follows a reasoning, ranking, and transfer pipeline to imbue dense models with both high performance and interpretability. First, STAR aligns a teacher LLM's reasoning with task objectives using a novel multi-granularity reward in Group Relative Policy Optimization (GRPO), leveraging only binary labels. Next, it refines the teacher's ability for calibrated scoring via token-level supervision, enabling efficient ranking through a single forward pass without any additional layers. Finally, this ''white-box'' knowledge is transferred to a compact student via multi-faceted distillation that preserves both reasoning logic and ranking behavior. Offline experiments demonstrate that our 0.6B student model rivals the performance of a strong 8B baseline, making it highly efficient and fully deployable. Real-world effectiveness is validated by significant online A/B test gains, including a +0.93% GoodRate lift and a +1.04% increase in GMV. STAR has been fully deployed to 100% of main search traffic on 1688.com.

In large-scale e-commerce search, accurately modeling multi-modal relevance is paramount for matching user intent—especially given the growing influence of visual content on shopping decisions. However, existing methods often fail to perform fine-grained reasoning. For instance, they struggle when a product title is irrelevant due to marketing language, but its image is highly relevant to the query. Furthermore, they cannot effectively disambiguate which specific query terms are satisfied by the visual versus the textual modality. While Large Language Models (LLMs) excel at such reasoning, their high computational overhead makes direct online deployment infeasible. To bridge this gap, we propose TRACE (Term-level Reasoning And Chain-of-thought Enhanced distillation), a framework designed for deploying advanced reasoning capabilities at scale. TRACE operates in two stages. First, it enhances an LLM's multi-modal reasoning by employing Group Relative Policy Optimization (GRPO) guided by a term-level Chain-of-Thought (CoT) reward function, enabling it to generate detailed, step-by-step relevance judgments. Second, it efficiently transfers this fine-grained reasoning to a lightweight, deployable model using a novel term-level knowledge distillation strategy that inherits reasoning ability. Offline evaluations show significant improvements across different datasets. More critically, online A/B tests on 1688.com resulted in a +1.04% GMV uplift, a +0.906% LTV increase, and a +0.523% improvement in UV_L2O, demonstrating its significant value in a real-world production environment.

Knowledge hypergraph representation learning, which projects entities and$n$-ary relations into a low-dimensional vector space, remains a challenging area to be explored despite the ubiquity of$n$-ary relational facts in the real world. Current methods are always extensions of those used for knowledge graphs with shallow or deep structures. However, shallow and linear models limit the extraction capacity of the latent knowledge, while deep and non-linear models lead to the overabundance of parameters. In this paper, we propose a novel knowledge hypergraph completion model called HJE, which utilizes the powerful capability of convolutional neural networks for efficient representation learning. Interaction-enhanced 3D convolution and relation-aware 2D convolution are jointly utilized by HJE to extract explicit and implicit global knowledge and semantic information effectively without compromising the translation property of the model. Moreover, HJE constructs a unified learnable embedding matrix to capture entity position information in knowledge tuples. The entity mask mechanism can naturally couple the multilinear scoring approach for$n$-ary facts to speed up the training convergence of the model. Extensive experimental results on real datasets of knowledge hypergraphs and knowledge graphs demonstrate the superior performance of HJE compared with state-of-the-art baselines.

Knowledge hypergraph embedding, which projects entities and n-ary relations into a low-dimensional continuous vector space to predict missing links, remains a challenging area to be explored despite the ubiquity of n-ary relational facts in the real world. Currently, knowledge hypergraph link prediction methods are essentially simple extensions of those used in knowledge graphs, where n-ary relational facts are decomposed into different subelements. Convolutional neural networks have been shown to have remarkable information extraction capabilities in previous work on knowledge graph link prediction. In this paper, we propose a novel embedding-based knowledge hypergraph link prediction model named HyConvE, which exploits the powerful learning ability of convolutional neural networks for effective link prediction. Specifically, we employ 3D convolution to capture the deep interactions of entities and relations to efficiently extract explicit and implicit knowledge in each n-ary relational fact without compromising its translation property. In addition, appropriate relation and position-aware filters are utilized sequentially to perform two-dimensional convolution operations to capture the intrinsic patterns and position information in each n-ary relation, respectively. Extensive experimental results on real datasets of knowledge hypergraphs and knowledge graphs demonstrate the superior performance of HyConvE compared with state-of-the-art baselines.

Abstract Link prediction in knowledge hypergraphs is essential for various knowledge-based applications, including question answering and recommendation systems. However, many current approaches simply extend binary relation methods from knowledge graphs to n-ary relations, which does not allow for capturing entity positional and role information in n-ary tuples. To address this issue, we introduce PosKHG, a method that considers entities’ positions and roles within n-ary tuples. PosKHG uses an embedding space with basis vectors to represent entities’ positional and role information through a linear combination, which allows for similar representations of entities with related roles and positions. Additionally, PosKHG employs a relation matrix to capture the compatibility of both information with all associated entities and a scoring function to measure the plausibility of tuples made up of entities with specific roles and positions. PosKHG achieves full expressiveness and high prediction efficiency. In experimental results, PosKHG achieved an average improvement of 4.1% on MRR compared to other state-of-the-art knowledge hypergraph embedding methods. Our code is available at https://anonymous.4open.science/r/PosKHG-C5B3/ .

Link prediction in knowledge hypergraphs has been recognized as a critical issue in various downstream tasks for knowledge-enabled applications, from question answering to recommender systems. However, most existing approaches are primarily performed in a black-box fashion, which learn low-dimensional embeddings for inference, thus cannot provide human-understandable interpretation. In this paper, we present HyperMLN, an n-ary, mixed, and explainable framework that interprets the path-reasoning process with first-order logic, which provides a knowledge-enhanced interpretable prediction framework, in which domain knowledge in the logic rules improves the performance of embedding models, while semantic information in the embedding space can optimize the weight of the logic rules in turn. To provide benchmark rule sets for explainable link prediction methods, three types of meta-logic rules in each popular dataset are mined for interpreting results. While achieving explainability, our framework also realizes an average improvement of 3.2% on [email protected] compared to the state-of-the-art knowledge hypergraph embedding method. Our code is available at https://github.com/zirui-chen/HyperMLN.