Generative Retrieval (GR) is rapidly transforming e-commerce search by replacing traditional multi-stage pipelines with the autoregressive decoding of structured Semantic IDs (SIDs). Despite this architectural efficiency, aligning GR models with nuanced, realworld user preferences remains a critical challenge. While Direct Preference Optimization (DPO) offers an efficient alignment solution, its direct application to structured SIDs suffers from three limitations: (i) it penalizes shared hierarchical prefixes, causing gradient conflicts; (ii) it is vulnerable to noisy pseudo-negatives from implicit feedback; and (iii) in multi-label queries with multiple relevant items, it exacerbates a probability "squeezing effect" among valid candidates. To address these issues, we propose RAD-DPO, which introduces token-level gradient detachment to protect prefix structures, similarity-based dynamic reward weighting to mitigate label noise, and a multi-label global contrastive objective integrated with global SFT loss to explicitly expand positive coverage. Extensive offline evaluations and large-scale online A/B testing on JD.com's core search engine demonstrate that RAD-DPO achieves significant improvements in both retrieval precision and training efficiency, proving its robustness for massive industrial deployments

Recommender systems aim to leverage user interaction signals to recommend items that users are likely to be interested in. Motivated by the success of Large Language Model (LLMs), generative recommendation (GR) has recently gained increasing attention, typically following a two-stage paradigm: supervised fine-tuning followed by preference alignment. However, aligning generative recommenders with users' personalized preferences remains challenging, as user feedback is inherently heterogeneous and uncertain. Different types of user interaction signals reflect varying levels of intent and should therefore be modeled differently. In this work, we propose Adaptive Mix Preference Optimization (AMPO), an adaptive alignment framework that mixes likelihood and preference objectives with self-calibrated, sample-wise confidence adjustment. AMPO introduces an adaptive target margin that leverages the model's own probability ratio to modulate optimization strength: confident pairs receive full margins that reinforce correct rankings, while uncertain pairs receive reduced margins that prevent overfitting to ambiguous signals. Additionally, AMPO incorporates negative log-likelihood regularization on preferred items to counteract likelihood displacement, a phenomenon where contrastive objectives cause preferred and non-preferred probabilities to collapse simultaneously. Such a design eliminates the need for a reference model, yielding up to 2.5x speedup and 35% memory reduction. Extensive experiments on public benchmarks and a large-scale industrial dataset demonstrate consistent improvements in ranking metrics. Online A/B tests on a major e-commerce platform further confirm statistically significant gains in click-through and conversion rates. The code is available at https://github.com/jumbo-q/ampo.

Novelty recommendation sustains long-term user engagement by exposing users to content that is both relevant and meaningfully different from their recent consumption. In large-scale e-commerce, this requires composing coherent yet non-redundant recommendation lists, a task fundamentally constrained by the relevance-diversity trade-off. Large language models (LLMs) offer a unified generative paradigm for inferring user intent and producing semantically coherent candidates, yet industrial deployment faces two critical challenges: (i) scarce supervision for modeling novelty transitions and diversity-aware list construction, and (ii) reward granularity mismatch, where standard RL assigns coarse sequence-level rewards that fail to capture item-level redundancy and complementarity. We present BALANCE, a hierarchical reasoning-and-generation framework that decomposes novelty recommendation into three structured stages: generating a Novelty Tag for exploration direction, refining an Interest Topic for intent specification, and constructing a Recommendation List for facet coverage. We address data scarcity through a self-reflection pipeline that synthesizes high-quality supervision by integrating real behavior logs with structured rationales. We resolve granularity mismatch through Sequence-Item Policy Optimization (SIPO), which jointly optimizes sequence- and item-level objectives via granularity-aware advantage fusion. Extensive offline experiments and online A/B test on the JD.com recommender system, validate the performance of our method, highlighting its superior novelty and diversity without compromising relevance.

Generative retrieval formulates document retrieval as an identifier generation task. While prevailing methods increasingly adopt Semantic IDs (SIDs), their opaque nature and rigid mappings struggle with the dynamic inventory and strict interpretability requirements of E-commerce search. Furthermore, generating accurate targets from brief queries against noisy, loosely structured item titles remains a practical challenge. To address these issues, we propose a Query-to-Multi-Span generative retrieval framework tailored for E-commerce. Instead of relying on opaque SIDs or raw titles, our method simplifies the process by generating interpretable multispan identifiers from queries. We align the autoregressive model with user preferences using click logs, and employ a constraintbased beam search to isolate key spans for final item retrieval. This approach explicitly bridges generative models with robust constraint matching, ensuring both matching accuracy and transparency. Extensive offline evaluations demonstrate competitive retrieval performance, and online A/B tests confirm its effectiveness in delivering measurable conversion gains in a production environment.

While semantic ID-based generative retrieval enables efficient end-to-end modeling in industrial applications, these methods face a persistent trade-off. On one hand, data-rich head items often suffer from ID collisions, which blur their distinct features and degrade downstream tasks. On the other hand, data-sparse tail items especially cold-start items are prone to semantic fragmentation during quantization; they are often mapped as isolated discrete points, which severely hinders their ability to generalize. To address this issue, we propose the Anchored Curriculum with Sequential Adaptive Quantization (SA2CRQ) framework. The framework introduces Sequential Adaptive Residual Quantization (SARQ) to dynamically allocate code lengths based on item path entropy, assigning longer, discriminative IDs to head items and shorter, generalizable IDs to tail items. To mitigate data sparsity, the Anchored Curriculum Residual Quantization (ACRQ) component utilizes a frozen semantic manifold learned from head items to regularize and accelerate the representation learning of tail items. Experimental results from a large-scale industrial search system and multiple public datasets indicate that SA2CRQ yields consistent improvements over existing baselines, particularly in cold-start retrieval scenarios.

User interest modeling plays a vital role in industrial recommendation systems (RSs). Existing generative recommendation (GR) methods rely on single-step direct inference, which falls short of deeply modeling complex and dynamically evolving user interest. Recent chain-of-thought (CoT)-based GR methods attempt to address this, but either suffer from information loss during semantic space transformation in explicit reasoning or yield uncontrollable, homogeneous reasoning chains in implicit reasoning.

Generative Retrieval (GR) offers a promising paradigm for recommendation through next-token prediction (NTP). However, scaling it to large-scale industrial systems introduces three challenges: (i) within a single request, the identical model inputs may produce inconsistent outputs due to the pagination request mechanism; (ii) the prohibitive cost of encoding long user behavior sequences with multi-token item representations based on semantic IDs, and (iii) aligning the generative policy with nuanced user preference signals. We present GenRec, a preference-oriented generative framework deployed on the JD App https://www.jd.com that addresses above challenges within a single decoder-only architecture. For training objective, we propose Page-wise NTP task, which supervises over an entire interaction page rather than each interacted item individually, providing denser gradient signal and resolving the one-to-many ambiguity of point-wise training. On the prefilling side, an asymmetric linear Token Merger compresses multi-token Semantic IDs in the prompt while preserving full-resolution decoding, reducing input length by ~2× with negligible accuracy loss. To further align outputs with user satisfaction, we introduce GRPO-SR, a reinforcement learning method that pairs Group Relative Policy Optimization with NLL regularization for training stability, and employs Hybrid Rewards combining a dense reward model with a relevance gate to mitigate reward hacking. In month-long online A/B tests serving production traffic, GenRec achieves 9.5% improvement in click count and 8.7% in transaction count over the existing pipeline.

State-of-the-art approaches for click-through rate (CTR) prediction in industry predominantly rely on transformer-based networks or their variants. However, as user behavior sequences become longer, employing self-attention networks for CTR prediction within a constrained inference time presents a significant challenge. To address this, mainstream methods adopt a classical two-stage paradigm: a General Search Unit (GSU) for quickly retrieving relevant items from long-term behaviors, and an Exact Search Unit (ESU) for applying effective Multi-Head Target Attention (MHTA) over the items selected by the GSU. These two-stage algorithms have certain limitations. Firstly, the GSU needs to retrieve different target subsequences for different target items, restricting the ESU to a suboptimal MHTA network rather than a more effective transformer-based network. Secondly, the GSU retrieves only a subset of items from the user's behavior sequence, ignoring the evolution of user interests and the interrelationships between different points of interest. To overcome these challenges, we propose a novel end-to-end hierarchical user long-term behavior modeling network for CTR prediction (HBM). Specifically, we employ the multi-interest routing layer to channel the user's long-term behavior to several aggregated interest clusters. Furthermore, we introduce a fine interest learning network that selects the top-k interests from the initial aggregated representations. Subsequently, we employ a transformer network to model the user's behavior sequence associated with these top-k interests in a detailed manner, while also capturing the inherent correlations between different user interests at a coarse level. Finally, we integrate the coarse and fine interests. Extensive experiments on two real-world datasets demonstrate the effectiveness of our proposed methods. In addition, an online A/B test on the JD recommendation platform shows promising improvements, with a 2.15% increase in CTR and a 0.98% increase in CVR, accompanied by lower online inference latency.

A key challenge in personalized product search is to capture user’s preferences. Recent work attempted to model sequences of user historical behaviors, i.e., product purchase histories, to build user profiles and to personalize results accordingly. Although these approaches have demonstrated promising retrieval performances, we notice that most of them focus solely on the intra-sequence interactions between items. However, as there is usually a small amount of historical behavior data, the user profiles learned by these approaches could be very sensitive to the noise included in it. To tackle this problem, we propose incorporating out-of-sequence external information to enhance user modeling. More specifically, we inject the external item–item relations (e.g., belonging to the same brand), and query–query relations (e.g., the semantic similarities between them), into the intra-sequence interaction to learn better user profiles. In addition, we devise two auxiliary decoders, with the historical item sequence reconstruction task and the global item similarity prediction task, to further improve the reliability of user modeling. Experimental results on two datasets from simulated and real user search logs respectively show that the proposed personalized product search method outperforms existing approaches.

Conversational Recommendation Systems (CRS) effectively address information asymmetry by dynamically eliciting user preferences through multi-turn interactions. However, existing CRS methods commonly assume that users have clear, definite preferences for one or multiple target items. This assumption can lead to over-trusting user feedback, treating accepts/rejects as definitive signals to filter items and reduce the candidate space, potentially causing over-filtering and excluding relevant alternatives. In reality, users often exhibit vague preferences, lacking well-defined inclinations for certain attribute types (e.g., color, pattern), and their decision-making process during interactions is rarely binary. Instead, users’ choices are relative, reflecting a range of preferences rather than strict likes or dislikes. To address this issue, we introduce a novel scenario called Vague Preference Multi-Round Conversational Recommendation (VPMCR), which employs a soft estimation mechanism to assign non-zero confidence scores to all candidate items, accommodating users’ vague and dynamic preferences while mitigating over-filtering. In the VPMCR setting, we introduce a solution called Vague Preference Policy Learning (VPPL), which consists of two main components: Ambiguity-Aware Soft Estimation (ASE) and Dynamism-Aware Policy Learning (DPL). ASE aims to accommodate the ambiguity in user preferences by estimating preference scores for both directed and inferred preferences, employing a choice-based approach and a time-aware preference decay strategy. DPL implements a policy learning framework, leveraging the preference distribution from ASE, to guide the conversation and adapt to changes in users’ preferences for making recommendations or querying attributes. Extensive experiments conducted on diverse datasets demonstrate the effectiveness of VPPL within the VPMCR framework, outperforming existing methods and setting a new benchmark for CRS research. Our work represents a significant advancement in accommodating the inherent ambiguity and relative decision-making processes exhibited by users, improving the overall performance and applicability of CRS in real-world settings.

Traffic allocation is a process of redistributing natural traffic to products by adjusting their positions in the post-search phase, aimed at effectively fostering merchant growth, precisely meeting customer demands, and ensuring the maximization of interests across various parties within e-commerce platforms. Existing methods based on learning to rank neglect the long-term value of traffic allocation, whereas approaches of reinforcement learning suffer from balancing multiple objectives and the difficulties of cold starts within real-world data environments. To address the aforementioned issues, this paper propose a multi-objective deep reinforcement learning framework consisting of multi-objective Q-learning (MOQ), a decision fusion algorithm (DFM) based on the cross-entropy method(CEM), and a progressive data augmentation system (PDA). Specifically. MOQ constructs ensemble RL models, each dedicated to an objective, such as click-through rate, conversion rate, etc. These models individually determine the position of items as actions, aiming to estimate the long-term value of multiple objectives from an individual perspective. Then we employ DFM to dynamically adjust weights among objectives to maximize long-term value, addressing temporal dynamics in objective preferences in e-commerce scenarios. Initially, PDA trained MOQ with simulated data from offline logs. As experiments progressed, it strategically integrated real user interaction data, ultimately replacing the simulated dataset to alleviate distributional shifts and the cold start problem. Experimental results on real-world online e-commerce systems demonstrate the significant improvements of MODRL-TA, and we have successfully deployed MODRL-TA on an e-commerce search platform.

In the domain of e-commerce, query rewriting is a potent strategy for bridging the lexical gap between search queries and product descriptions, thereby enhancing the recall rate of search engines. This research introduces a query rewriting framework predicated on large language models (LLM), encompassing three phases of training: domain-specific pre-training, supervised fine-tuning (SFT) and reinforcement learning (RL) for objective alignment. To detail, the process initiates with domain-specific pre-training using consumer behavior data and product descriptions from JD.com. Subsequently, we filter and utilize high-quality query-rewrite pairs for SFT. The final stage employs RL to refine the model's objective alignment, utilizing an offline search system as the simulation environment. The RL's training reward is derived from the recall rate, aiming to optimize the number of relevant products the rewrites retrieve. Through offline evaluations, our method has demonstrated its capacity to substantially enhance the efficacy of LLMs for e-commerce query rewriting. Moreover, online A/B testing has corroborated that our approach significantly boosts the number of purchases made per user (UCVR). Since December 2023, our approach has been successfully implemented on JD.com, one of China's most frequented online shopping platforms.

In the rapidly evolving field of e-commerce, the effectiveness of search re-ranking models is crucial for enhancing user experience and driving conversion rates. Despite significant advancements in feature representation and model architecture, the integration of multimodal information remains underexplored. This study addresses this gap by investigating the computation and fusion of textual and visual information in the context of re-ranking. We propose Advancing Re-ranking with Multimodal Fusion and Target-Oriented Auxiliary Tasks (ARMMT), which integrates an attention-based multimodal fusion technique and an auxiliary ranking-aligned task to enhance item representation and improve targeting capabilities. This method not only enriches the understanding of product attributes but also enables more precise and personalized recommendations. Experimental evaluations on JD.com's search platform demonstrate that ARMMT achieves state-of-the-art performance in multimodal information integration, evidenced by a 0.22% increase in the Conversion Rate (CVR), significantly contributing to Gross Merchandise Volume (GMV). This pioneering approach has the potential to revolutionize e-commerce re-ranking, leading to elevated user satisfaction and business growth.

Multi-agent collaboration is the latest trending method to build conversational recommender systems (CRS), especially with the widespread use of Large Language Models (LLMs) recently. Typically, these systems employ several LLM agents, each serving distinct roles to meet user needs. In an industrial setting, it’s essential for a CRS to exhibit low first token latency (i.e., the time taken from a user’s input until the system outputs its first response token.) and high scalability—for instance, minimizing the number of LLM inferences per user request—to enhance user experience and boost platform profit. For example, JD.com’s baseline CRS features two LLM agents and a search API but suffers from high first token latency and requires two LLM inferences per request (LIPR), hindering its performance. To address these issues, we introduce a Hybrid Multi-Agent Collaborative Recommender System (Hybrid-MACRS). It includes a central agent powered by a fine-tuned proprietary LLM and a search agent combining a related search module with a search engine. This hybrid system notably reduces first token latency by about 70% and cuts the LIPR from 2 to 1. We conducted thorough online A/B testing to confirm this approach’s efficiency.

Search and recommendation (S&R) are the two most important scenarios in e-commerce. The majority of users typically interact with products in S&R scenarios, indicating the need and potential for joint modeling. Traditional multi-scenario models use shared parameters to learn the similarity of multiple tasks, and task-specific parameters to learn the divergence of individual tasks. This coarse-grained modeling approach does not effectively capture the differences between S&R scenarios. Furthermore, this approach does not sufficiently exploit the information across the global label space. These issues can result in the suboptimal performance of multi-scenario models in handling both S&R scenarios. To address these issues, we propose an effective and universal framework for Unified Search and Recommendation (USR), designed with S&R Views User Interest Extractor Layer (IE) and S&R Views Feature Generator Layer (FG) to separately generate user interests and scenario-agnostic feature representations for S&R. Next, we introduce a Global Label Space Multi-Task Layer (GLMT) that uses global labels as supervised signals of auxiliary tasks and jointly models the main task and auxiliary tasks using conditional probability. Extensive experimental evaluations on real-world industrial datasets show that USR can be applied to various multi-scenario models and significantly improve their performance. Online A/B testing also indicates substantial performance gains across multiple metrics. Currently, USR has been successfully deployed in the 7Fresh App.

Re-ranking is a process of rearranging ranking list to more effectively meet user demands by accounting for the interrelationships between items. Existing methods predominantly enhance the precision of search results, often at the expense of diversity, leading to outcomes that may not fulfill the varied needs of users. Conversely, methods designed to promote diversity might compromise the precision of the results, failing to satisfy the users' requirements for accuracy. To alleviate the above problems, this paper proposes a Preference-oriented Diversity Model Based on Mutual-information (PODM-MI), which consider both accuracy and diversity in the re-ranking process. Specifically, PODM-MI adopts Multidimensional Gaussian distributions based on variational inference to capture users' diversity preferences with uncertainty. Then we maximize the mutual information between the diversity preferences of the users and the candidate items using the maximum variational inference lower bound to enhance their correlations. Subsequently, we derive a utility matrix based on the correlations, enabling the adaptive ranking of items in line with user preferences and establishing a balance between the aforementioned objectives. Experimental results on real-world online e-commerce systems demonstrate the significant improvements of PODM-MI, and we have successfully deployed PODM-MI on an e-commerce search platform.

Click-Through Rate (CTR) prediction plays a pivotal role in numerous industrial applications, including online advertising and recommender systems. Existing approaches primarily focus on modeling the correlation between user interests and candidate items. However, we argue that personalized user preferences for candidate items depend not only on correlation but also on the satisfaction of associated interests. To address this limitation, we propose SUIN, a novel CTR model that integrates satisfaction factors into user interest modeling for enhanced click-through rate prediction. Specifically, we employ a user interest satisfaction-aware network to capture the degree of satisfaction for each interest, thereby enabling adaptation of the user's personalized preference based on satisfaction levels. Additionally, we leverage the exposure-unclicked signal (recommended to the user but not clicked) as supervision during training, facilitating the interest satisfaction module to better model the satisfaction degree of user interests. Besides, this module serves as a foundational building block suitable for integration into mainstream sequential-based CTR models. Extensive experiments conducted on two real-world datasets demonstrate the superiority of our proposed model, outperforming state-of-the-art methods across various evaluation metrics. Furthermore, an online A/B test deployed on large-scale recommender systems shows significant improvements achieved by our model in diverse evaluation metrics.

Click-Through Rate (CTR) prediction is becoming increasingly vital in many industrial applications, such as recommendations and online advertising. How to precisely capture users' dynamic and evolving interests from previous interactions (e.g., clicks, purchases, etc.) is a challenging task in CTR prediction. Mainstream approaches focus on disentangling user interests in a heuristic way or modeling user interests into a static representation. However, these approaches overlook the importance of users' current intent and the complex interactions between their current intent and global interests. To address these concerns, in this paper, we propose a novel intent-enhanced user interest modeling for click-through rate prediction in large-scale e-commerce recommendations, abbreviated as IUI. Methodologically, different from existing works, we consider users' recent interactions to be inspired by their implicit intent and then leverage an intent-aware network to model their current local interests in a more precise and fine-grained manner. In addition, to obtain a more stable co-dependent global and local interest representation, we employ a co-attention network capable of activating the corresponding interest in global-level interactions and capturing the dynamic interactions between global- and local-level interaction behaviors. Finally, we incorporate self-supervised learning into the model training by maximizing the mutual information between the global and local representations obtained via the above two networks to enhance the CTR prediction performance. Compared with existing methods, IUI benefits from the different granularity of user interest to generate a more accurate and comprehensive preference representation. Experimental results demonstrate that the proposed model outperforms previous state-of-the-art methods in various metrics on three real-world datasets. In addition, an online A/B test deployed on the JD recommendation platforms shows a promising improvement across multiple evaluation metrics.

Ranking model plays an essential role in e-commerce search and recommendation. An effective ranking model should give a personalized ranking list for each user according to the user preference. Existing algorithms usually extract a user representation vector from the user behavior sequence, then feed the vector into a feed-forward network (FFN) together with other features for feature interactions, and finally produce a personalized ranking score. Despite tremendous progress in the past, there is still room for improvement. Firstly, the personalized patterns of feature interactions for different users are not explicitly modeled. Secondly, most of existing algorithms have poor personalized ranking results for long-tail users with few historical behaviors due to the data sparsity.To overcome the two challenges, we propose Attention Weighted Mixture of Experts (AW-MoE) with contrastive learning for personalized ranking. Firstly, AW-MoE leverages the MoE framework to capture personalized feature interactions for different users. To model the user preference, the user behavior sequence is simultaneously fed into expert networks and the gate network. Within the gate network, one gate unit and one activation unit are designed to adaptively learn the fine-grained activation vector for experts using an attention mechanism. Secondly, a random masking strategy is applied to the user behavior sequence to simulate long-tail users, and an auxiliary contrastive loss is imposed to the output of the gate network to improve the model generalization for these users. This is validated by a higher performance gain on the long-tail user test set.Experiment results on a JD real production dataset and a public dataset demonstrate the effectiveness of AW-MoE, which significantly outperforms state-of-art methods. Notably, AW-MoE has been successfully deployed in the JD e-commerce search engine, serving the real traffic of hundreds of millions of active users.

Classical recommendation methods typically face the filter bubble problem where users likely receive recommendations of their familiar items, making them bored and dissatisfied. To alleviate such an issue, this applied paper introduces a novel framework for personalized serendipitous recommendation in an e-commerce platform (i.e., JD.com), which allows to present user unexpected and satisfying items deviating from user’s prior behaviors, considering both accuracy and novelty. To achieve such a goal, it is crucial yet challenging to recognize when a user is willing to receive serendipitous items and how many novel items are expected. To address above two challenges, a two-stage framework is designed. Firstly, a DNN-based scorer is deployed to quantify the novelty degree of a product category based on user behavior history. Then, we resort to a potential outcome framework to decide the optimal timing to recommend a user serendipitous items and the novelty degree of the recommendation. Online A/B test on the e-commerce recommender platform in JD.com demonstrates that our model achieves significant gains on various metrics, 0.54% relative increase of impressive depth, 0.8% of average user click count, 3.23% and 1.38% of number of novel impressive and clicked items individually.

The embedding index has become an essential part of the dense retrieval (DR) system, which enables a fast search for billion of items in online E-commerce applications. To accelerate the retrieval process in industrial scenarios, most of the previous studies only utilize item embeddings. However, the product quantization process without query embeddings will lead to inconsistency between queries and items. A straightforward solution is to put query embedding into the product quantization process. But we found that the distance of the positive query and item embedding pairs is too large, which means the query and item embeddings learned by the two-tower are not fully aligned. This problem would lead to performance decay when directly putting query embeddings into the product quantization.

Recently, personalized product search attracts great attention and many models have been proposed. To evaluate the effectiveness of these models, previous studies mainly utilize the simulated Amazon recommendation dataset, which contains automatically generated queries and excludes cold users and tail products. We argue that evaluating with such a dataset may yield unreliable results and conclusions, and deviate from real user satisfaction. To overcome these problems, in this paper, we release a personalized product search dataset comprised of real user queries and diverse user-product interaction types (clicking, adding to cart, following, and purchasing) collected from JD.com, a popular Chinese online shopping platform. More specifically, we sample about 170,000 active users on a specific date, then record all their interacted products and issued queries in one year, without removing any tail users and products. This finally results in roughly 12,000,000 products, 9,400,000 real searches, and 26,000,000 user-product interactions. We study the characteristics of this dataset from various perspectives and evaluate representative personalization models to verify its feasibility. The dataset can be publicly accessed at Github: https://github.com/rucliujn/JDsearch.

Semantic retrieval, which retrieves semantically matched items given a textual query, has been an essential component to enhance system effectiveness in e-commerce search. In this paper, we study the multimodal retrieval problem, where the visual information (e.g, image) of item is leveraged as supplementary of textual information to enrich item representation and further improve retrieval performance. Though learning from cross-modality data has been studied extensively in tasks such as visual question answering or media summarization, multimodal retrieval remains a non-trivial and unsolved problem especially in the asymmetric scenario where the query is unimodal while the item is multimodal. In this paper, we propose a novel model named SMAR, which stands for Semantic-enhanced Modality-Asymmetric Retrieval, to tackle the problem of modality fusion and alignment in this kind of asymmetric scenario. Extensive experimental results on an industrial dataset show that the proposed model outperforms baseline models significantly in retrieval accuracy. We have open sourced our industrial dataset for the sake of reproducibility and future research works.

BERT-style models pre-trained on the general corpus (e.g., Wikipedia) and fine-tuned on specific task corpus, have recently emerged as breakthrough techniques in many NLP tasks: question answering, text classification, sequence labeling and so on. However, this tech- nique may not always work, especially for two scenarios: a corpus that contains very different text from the general corpus Wikipedia, or a task that learns embedding spacial distribution for a specific purpose (e.g., approximate nearest neighbor search). In this paper, to tackle the above two scenarios that we have encountered in an industrial e-commerce search system, we propose customized and novel pre-training tasks for two critical modules: user intent detec- tion and semantic embedding retrieval. The customized pre-trained models after fine-tuning, being less than 10% of BERT-base's size in order to be feasible for cost-efficient CPU serving, significantly improve the other baseline models: 1) no pre-training model and 2) fine-tuned model from the official pre-trained BERT using general corpus, on both offline datasets and online system. We have open sourced our datasets 1 for the sake of reproducibility and future works.

Recent years have seen a significant amount of interests in Sequential Recommendation (SR), which aims to understand and model the sequential user behaviors and the interactions between users and items over time. Surprisingly, despite the huge success Sequential Recommendation has achieved, there is little study on Sequential Search (SS), a twin learning task that takes into account a user's current and past search queries, in addition to behavior on historical query sessions. The SS learning task is even more important than the counterpart SR task for most of E-commence companies due to its much larger online serving demands as well as traffic volume.

Nowadays e-commerce search has become an integral part of many people's shopping routines. One critical challenge in today's e-commerce search is the semantic matching problem where the relevant items may not contain the exact terms in the user query. In this paper, we propose a novel deep neural network based approach to query rewriting, in order to tackle this problem. Specifically, we formulate query rewriting into a cyclic machine translation problem to leverage abundant click log data. Then we introduce a novel cyclic consistent training algorithm in conjunction with state-of-the-art machine translation models to achieve the optimal performance in terms of query rewriting accuracy. In order to make it practical in industrial scenarios, we optimize the syntax tree construction to reduce computational cost and online serving latency. Offline experiments show that the proposed method is able to rewrite hard user queries into more standard queries that are more appropriate for the inverted index to retrieve. Comparing with human curated rule-based method, the proposed model significantly improves query rewriting diversity while maintaining good relevancy. Online A/B experiments show that it improves core e-commerce business metrics significantly. Since the summer of 2020, the proposed model has been launched into our search engine production, serving hundreds of millions of users.

Product search is the most common way for people to satisfy their shopping needs on e-commerce websites. Products are typically annotated with one of several broad categorical tags, such as "Clothing" or "Electronics", as well as finer-grained categories like "Refrigerator" or "TV", both under "Electronics". These tags are used to construct a hierarchy of query categories. Distributions of features such as price and brand popularity vary wildly across query categories. In addition, feature importance for the purpose of CTR/CVR predictions differs from one category to another. In this work, we leverage the Mixture of Expert (MoE) framework to learn a ranking model that specializes for each query category. In particular, our gate network relies solely on the category ids extracted from the user query.While classical MoE's pick expert towers spontaneously for each input example, we explore two techniques to establish more explicit and transparent connections between the experts and query categories. To help differentiate experts on their domain specialties, we introduce a form of adversarial regularization among the expert outputs, forcing them to disagree with one another. As a result, they tend to approach each prediction problem from different angles, rather than copying one another. This is validated by a much stronger clustering effect of the gate output vectors under different categories. In addition, soft gating constraints based on the categorical hierarchy are imposed to help similar products choose similar gate values. and make them more likely to share similar experts. This allows aggregation of training data among smaller sibling categories to overcome data scarcity.Experiments on a learning-to-rank dataset collected from the JD e-commerce search log demonstrate that MoE with these improvements consistently outperforms competing models, in terms of offline metrics and online AB tests.

Graph convolution networks (GCN), which recently becomes new state-of-the-art method for graph node classification, recommendation and other applications, has not been successfully applied to industrial-scale search engine yet. In this proposal, we introduce our approach, namely SearchGCN, for embedding-based candidate retrieval in one of the largest e-commerce search engine in the world. Empirical studies demonstrate that SearchGCN learns better embedding representations than existing methods, especially for long tail queries and items. Thus, SearchGCN has been deployed into JD.com's search production since July 2020.

Embedding index that enables fast approximate nearest neighbor(ANN) search, serves as an indispensable component for state-of-the-art deep retrieval systems. Traditional approaches, often separating the two steps of embedding learning and index building, incur additional indexing time and decayed retrieval accuracy. In this paper, we propose a novel method called Poeem, which stands for product quantization based embedding index jointly trained with deep retrieval model, to unify the two separate steps within an end-to-end training, by utilizing a few techniques including the gradient straight-through estimator, warm start strategy, optimal space decomposition and Givens rotation. Extensive experimental results show that the proposed method not only improves retrieval accuracy significantly but also reduces the indexing time to almost none. We have open sourced our approach for the sake of comparison and reproducibility.

Relevance has significant impact on user experience and business profit for e-commerce search platform. In this work, we propose a data-driven framework for search relevance prediction, by distilling knowledge from BERT and related multi-layer Transformer teacher models into simple feed-forward networks with large amount of unlabeled data. The distillation process produces a student model that recovers more than 97% test accuracy of teacher models on new queries, at a serving cost that's several magnitude lower (latency 150x lower than BERT-Base and 15x lower than the most efficient BERT variant, TinyBERT). The applications of temperature rescaling and teacher model stacking further boost model accuracy, without increasing the student model complexity. We present experimental results on both in-house e-commerce search relevance data as well as a public data set on sentiment analysis from the GLUE benchmark. The latter takes advantage of another related public data set of much larger scale, while disregarding its potentially noisy labels. Embedding analysis and case study on the in-house data further highlight the strength of the resulting model. By making the data processing and model training source code public, we hope the techniques presented here can help reduce energy consumption of the state of the art Transformer models and also level the playing field for small organizations lacking access to cutting edge machine learning hardwares.

As one of the largest B2C e-commerce platforms in China, JD.com also powers a leading advertising system, serving millions of advertisers with fingertip connection to hundreds of millions of customers. In our system, as well as most e-commerce scenarios, ads are displayed with images. This makes visual-aware Click Through Rate (CTR) prediction of crucial importance to both business effectiveness and user experience. Existing algorithms usually extract visual features using off-the-shelf Convolutional Neural Networks (CNNs) and late fuse the visual and non-visual features for the finally predicted CTR. Despite being extensively studied, this field still face two key challenges. First, although encouraging progress has been made in offline studies, applying CNNs in real systems remains non-trivial, due to the strict requirements for efficient end-to-end training and low-latency online serving. Second, the off-the-shelf CNNs and late fusion architectures are suboptimal. Specifically, off-the-shelf CNNs were designed for classification thus never take categories as input features. While in e-commerce, categories are precisely labeled and contain abundant visual priors that will help the visual modeling. Unaware of the ad category, these CNNs may extract some unnecessary category-unrelated features, wasting CNN's limited expression ability. To overcome the two challenges, we propose Category-specific CNN (CSCNN) specially for CTR prediction. CSCNN early incorporates the category knowledge with a light-weighted attention-module on each convolutional layer. This enables CSCNN to extract expressive category-specific visual patterns that benefit the CTR prediction. Offline experiments on benchmark and a 10 billion scale real production dataset from JD, together with an Online A/B test show that CSCNN outperforms all compared state-of-the-art algorithms. We also build a highly efficient infrastructure to accomplish end-to-end training with CNN on the 10 billion scale real production dataset within 24 hours, and meet the low latency requirements of online system (20ms on CPU). CSCNN is now deployed in the search advertising system of JD, serving the main traffic of hundreds of millions of active users.