Product retrieval is the backbone of e-commerce search: for each user query, it identifies a high-recall candidate set from billions of items, laying the foundation for high-quality ranking and user experience. Despite extensive optimization for mainstream queries, existing systems still struggle with long-tail queries, especially knowledge-intensive ones. These queries exhibit diverse linguistic patterns, often lack explicit purchase intent, and require domain-specific knowledge reasoning for accurate interpretation. They also suffer from a shortage of reliable behavioral logs, which makes such queries a persistent challenge for retrieval optimization.

Large Language Models (LLMs) are equipped with profound semantic knowledge, making them a natural choice for injecting semantic generalization into personalized search systems. However, in practice we find that directly fine-tuning LLMs on industrial personalized tasks (e.g. next item prediction) often yields suboptimal results. We attribute this bottleneck to a critical Knowledge--Action Gap: the inherent conflict between preserving pre-trained semantic knowledge and aligning with specific personalized actions by discriminative objectives. Empirically, action-only training objectives induce Semantic Collapse, such as attention ''sinks''. This degradation severely cripples the LLM's generalization, failing to bring improvements to personalized search systems.

Accurately estimating query-item relevance is vital for e-commerce ranking and conversion. While Large Language Models (LLMs) excel at reasoning, they often lack specialized knowledge required for long-tail or fast-evolving queries, necessitating Retrieval-Augmented Generation (RAG). However, production environments face three critical challenges: (1) external context is inherently noisy and inconsistent; (2) extreme latency budgets prohibit multi-stage processing or refinement; and (3) the model must simultaneously assess relevance and context-trust within a unified inference pass. We propose DyKnow-RAG, a reinforcement learning framework that teaches LLMs to learn to trust through dynamic utilization of external knowledge. Built on Group Relative Policy Optimization (GRPO), DyKnow-RAG utilizes a dual-group rollout strategy (parametric-only vs. with-context) and a posterior-driven inter-group advantage scaling mechanism. This enables the model to optimize context utilization without human process labels or extra inference overhead. Our pipeline further integrates structured Chain-of-Thought (CoT) and an uncertainty-prioritized RL pool to stabilize training. Offline evaluations show significant Macro-F1 and Accuracy gains, particularly on noise-sensitive query slices. Importantly, DyKnow-RAG has been deployed in Taobao's production system, serving hundreds of millions of active users and billions of daily search requests. Controlled A/B tests demonstrate consistent lifts in key business metrics, including GSB and Item Goodrate, while maintaining a p99 latency under 400ms. This work provides a scalable and deployable paradigm for operationalizing noisy RAG under extreme efficiency constraints of large-scale industrial search.

Query-product relevance prediction is fundamental to e-commerce search and has become even more critical in the era of AI-powered shopping, where semantic understanding and complex reasoning directly shape the user experience and business conversion. Large Language Models (LLMs) enable generative, reasoning-based approaches, typically aligned via supervised fine-tuning (SFT) or preference optimization methods like Direct Preference Optimization (DPO). However, the increasing complexity of business rules and user queries exposes the inability of existing methods to endow models with robust reasoning capacity for long-tail and challenging cases. Efforts to address this via reinforcement learning strategies like Group Relative Policy Optimization (GRPO) often suffer from sparse terminal rewards, offering insufficient guidance for multi-step reasoning and slowing convergence. To address these challenges, we propose TaoSR-AGRL, an Adaptive Guided Reinforcement Learning framework for LLM-based relevance prediction in Taobao Search Relevance. TaoSR-AGRL introduces two key innovations: (1) Rule-aware Reward Shaping, which decomposes the final relevance judgment into dense, structured rewards aligned with domain-specific relevance criteria; and (2) Adaptive Guided Replay, which identifies low-accuracy rollouts during training and injects targeted ground-truth guidance to steer the policy away from stagnant, rule-violating reasoning patterns toward compliant trajectories. TaoSR-AGRL was evaluated on large-scale real-world datasets and through online side-by-side human evaluations on Taobao Search. It consistently outperforms DPO and standard GRPO baselines in offline experiments, improving relevance accuracy, rule adherence, and training stability. The model trained with TaoSR-AGRL has been successfully deployed in the main search scenario on Taobao, serving hundreds of millions of users.

Cross-domain recommendation (CDR) aims to leverage the rich information from the source domain to enhance recommendation performance in the target domain. However, the data imbalance problem inherent across different domains compromises the effectiveness of CDR approaches, posing a significant challenge to CDR. Most current CDR methodologies focus on creating better user embeddings for the target domain, yet usually neglect the inconsistency in user activities due to data imbalance. As a result, the process of creating user embeddings tends to prioritize users with more frequent interactions and leave less active users underserved, leading these CDR methods to struggle in making accurate recommendations for those with fewer interactions. Such bias in creating embeddings reveals the fact that ''not all embeddings are created equal'' in CDR, which serves as the primary motivation of this study. Inspired by the recent development of contrastive learning, this paper proposes User-aware Contrastive Learning for Robust cross-domain recommendation (UCLR), enhancing the robustness of cross-domain recommendation. Specifically, our proposed method consists of two sub-modules: (i) pretrained global embedding, where the global user embeddings are pretrained across all the domains; (ii) contrastive dual-stream collaborative autoencoder, where more equal user embeddings are generated by optimizing contrastive loss with individualized temperatures. To further improve the performance of our method in each domain, we finetune the whole framework of UCLR based on Low-Rank Adaptation (LoRA). Theoretically, our method is equipped with a provable convergence guarantee during the contrastive learning stage. Furthermore, we also conduct comprehensive experiments on real-world datasets to validate the effectiveness of our proposed method.

Query classification plays a crucial role in e-commerce, where the goal is to assign user queries to appropriate categories within a hierarchical product category taxonomy. However, existing methods rely on a limited number of words from the category description and often neglect the hierarchical structure of the category tree, resulting in suboptimal category representations. To overcome these limitations, we propose a novel approach named hierarchical contrastive learning framework for query classification (HCL4QC), which leverages the hierarchical category tree structure to improve the performance of query classification. Specifically, HCL4QC is designed as a plugin module that consists of two innovative losses, namely local hierarchical contrastive loss (LHCL) and global hierarchical contrastive loss (GHCL). LHCL adjusts representations of categories according to their positional relationship in the hierarchical tree, while GHCL ensures the semantic consistency between the parent category and its child categories. Our proposed method can be adapted to any query classification tasks that involve a hierarchical category structure. We conduct experiments on two real-world datasets to demonstrate the superiority of our hierarchical contrastive learning. The results demonstrate significant improvements in the query classification task, particularly for long-tail categories with sparse supervised information.

Image retrieval system (IRS) is commonly used in E-Commerce platforms for a wide range of applications such as price comparison and commodity recommendation. However, customers may experience inconsistent retrieval problems. Although the retrieved image contains the query object, the main product of the retrieved image is not associated with the query product. This is caused by the wrong product instance location when building the product image retrieval library. We can easily determine which product is on sale through the hint of the title, so we propose Text-Guided MuliModal Product Location (TMML) to use additional product titles to assist in locating the actual selling product instance. We design a weakly-aligned region-text data collection method to generate region-text pseudo-label by utilizing the IRS and user behavior from the E-commerce platform. To mitigate the impact of data noise, we propose a Mutual-Aware Contrastive Loss. Our results show that the proposed TMML outperforms the state-of-the-art method GLIP [11] by 3.95% in top-1 precision on our multi-objects test set, and 2.53% error located images in AliExpress has been corrected, which greatly alleviates the retrieval inconsistencies in IRS.

Recommender systems have become a fundamental service in most E-Commerce platforms, in which the matching stage aims to retrieve potentially relevant candidate items to users for further ranking. Recently, some efforts on extracting multi-interests from user's historical behaviors have demonstrated superior performance. However, the historical behaviors are not noise-free due to the possible misclicks or disturbances. Existing works mainly overlook the fact that the interests of a user are not only reflected by the historical behaviors, but also inherently regulated by the profile information. Hence, we are interested in exploiting the benefit of user profile in multi-interest learning to enhance candidate matching performance. To this end, a user-aware multi-interest learning framework (named UMI) is proposed in this paper to exploit both user profile and behavior information for candidate matching. Specifically, UMI consists of two main components: dual-attention routing and interest refinement. In the dual-attention routing, we firstly introduce a user-guided attention network to identify the important historical items with respect to the user profile. Then, the resultant importance weights are leveraged via the dual-attentive capsule network to extract the user's multi-interests. Afterwards, the extracted interests are utilized to highlight the corresponding user profile features for interest refinement, such that different user profiles can be incorporated into interest learning for diverse user preference understanding. Besides, to improve the model's discriminative capacity, we further devise a harder-negatives strategy to support model optimization. Extensive experiments show that UMI significantly outperforms state-of-the-art multi-interest modeling alternatives. Currently, UMI has been successfully deployed at Taobao App in Alibaba, serving hundreds of millions of users.

The micro-video recommendation system becomes an essential part of the e-commerce platform, which helps disseminate micro-videos to potentially interested users. Existing micro-video recommendation methods only focus on users' browsing behaviors on micro-videos, but ignore their purchasing intentions in the e-commerce environment. Thus, they usually achieve unsatisfied e-commerce micro-video recommendation performances. To address this problem, we design a sequential multi-modal information transfer network (SEMI), which utilizes product-domain user behaviors to assist micro-video recommendations. SEMI effectively selects relevant items (i.e., micro-videos and products) with multi-modal features in the micro-video domain and product domain to characterize users' preferences. Moreover, we also propose a cross-domain contrastive learning (CCL) algorithm to pre-train sequence encoders for modeling users' sequential behaviors in these two domains. The objective of CCL is to maximize a lower bound of the mutual information between different domains. We have performed extensive experiments on a large-scale dataset collected from Taobao, a world-leading e-commerce platform. Experimental results show that the proposed method achieves significant improvements over state-of-the-art recommendation methods. Moreover, the proposed method has also been deployed on Taobao, and the online A/B testing results further demonstrate its practical value.

The large-scale recommender system mainly consists of two stages: matching and ranking. The matching stage (also known as the retrieval step) identifies a small fraction of relevant items from billion-scale item corpus in low latency and computational cost. Item-to-item collaborative filtering (item-based CF) and embedding-based retrieval (EBR) have been long used in the industrial matching stage owing to its efficiency. However, item-based CF is hard to meet personalization, while EBR has difficulty in satisfying diversity. In this paper, we propose a novel matching architecture, Path-based Deep Network (named PDN), through incorporating both personalization and diversity to enhance matching performance. Specifically, PDN is comprised of two modules: Trigger Net and Similarity Net. PDN utilizes Trigger Net to capture the user's interest in each of his/her interacted item. Similarity Net is devised to evaluate the similarity between each interacted item and the target item based on these items' profile and CF information. The final relevance between the user and the target item is calculated by explicitly considering user's diverse interests, \ie aggregating the relevance weights of the related two-hop paths (one hop of a path corresponds to user-item interaction and the other to item-item relevance). Furthermore, we describe the architecture design of the proposed PDN in a leading real-world E-Commerce service (Mobile Taobao App). Based on offline evaluations and online A/B test, we show that PDN outperforms the existing solutions for the same task. The online results also demonstrate that PDN can retrieve more personalized and more diverse items to significantly improve user engagement. Currently, PDN system has been successfully deployed at Mobile Taobao App and handling major online traffic.

Domain-specific knowledge bases play an increasingly important role in a variety of real applications. In this paper, we use the product knowledge base in the largest Chinese e-commerce platform, Taobao, as an example to investigate a completion procedure of a domain-specific knowledge base. We argue that the domain-specific knowledge bases tend to be incomplete, and are oblivious to their incompleteness, without a continuous completion procedure in place. The key component of this completion procedure is the classification of emerging query terms into corresponding properties of categories in existing taxonomy. Our proposal is that we use query logs to complete the product knowledge base of Taobao. However, the query driven completion usually faces many challenges including distinguishing the fine-grained semantic of unrecognized terms, handling the sparse data and so on. We propose a graph based solution to overcome these challenges. We first construct a lot of positive evidence to establish the semantical similarity between terms, and then run a shortest path or alternatively a random walk on the similarity graph under a set of constraints derived from a set of negative evidence to find the best candidate property for emerging query terms. We finally conduct extensive experiments on real data of Taobao and a subset of CN-DBpedia. The results show that our solution classifies emerging query terms with a good performance. Our solution is already deployed in Taobao, helping it find nearly 7 million new values for properties. The complete product knowledge base significantly improves the ratio of recognized queries and recognized terms by more than 25% and 32%, respectively.

Relevance is the core problem of a search engine, and one of the main challenges is the vocabulary gap between user queries and documents. This problem is more serious in e-commerce, because language in product titles is more professional. Query rewriting and semantic matching are two key techniques to bridge the semantic gap between them to improve relevance. Recently, deep neural networks have been successfully applied to the two tasks and enhanced the relevance performance. However, such approaches suffer from the sparseness of training data in e-commerce scenario. In this study, we investigate the instinctive connection between query rewriting and semantic matching tasks, and propose a co-training framework to address the data sparseness problem when training deep neural networks. We first build a huge unlabeled dataset from search logs, on which the two tasks can be considered as two different views of the relevance problem. Then we iteratively co-train them via labeled data generated from this unlabeled set to boost their performance simultaneously. We conduct a series of offline and online experiments on a real-world e-commerce search engine, and the results demonstrate that the proposed method improves relevance significantly.

In this paper, we investigate the task of aggregating search results from heterogeneous sources in an E-commerce environment. First, unlike traditional aggregated web search that merely presents multi-sourced results in the first page, this new task may present aggregated results in all pages and has to dynamically decide which source should be presented in the current page. Second, as pointed out by many existing studies, it is not trivial to rank items from heterogeneous sources because the relevance scores from different source systems are not directly comparable. To address these two issues, we decompose the task into two subtasks in a hierarchical structure: a high-level task for source selection where we model the sequential patterns of user behaviors onto aggregated results in different pages so as to understand user intents and select the relevant sources properly; and a low-level task for item presentation where we formulate a slot filling process to sequentially present the items instead of giving each item a relevance score when deciding the presentation order of heterogeneous items. Since both subtasks can be naturally formulated as sequential decision problems and learn from the future user feedback on search results, we build our model with hierarchical reinforcement learning. Extensive experiments demonstrate that our model obtains remarkable improvements in search performance metrics, and achieves a higher user satisfaction.