Generative retrieval introduces a groundbreaking paradigm to document retrieval by directly generating the identifier of a pertinent document in response to a specific query. This paradigm has demonstrated considerable benefits and potential, particularly in representation and generalization capabilities, within the context of large language models. However, it faces significant challenges in E-commerce search scenarios, including the complexity of generating detailed item titles from brief queries, the presence of noise in item titles with weak language order, issues with long-tail queries, and the interpretability of results. To address these challenges, we have developed an innovative framework for E-commerce search, called generative retrieval with preference optimization. This framework is designed to effectively learn and align an autoregressive model with target data, subsequently generating the final item through constraint-based beam search. By employing multi-span identifiers to represent raw item titles and transforming the task of generating titles from queries into the task of generating multi-span identifiers from queries, we aim to simplify the generation process. The framework further aligns with human preferences using click data and employs a constrained search method to identify key spans for retrieving the final item, thereby enhancing result interpretability. Our extensive experiments show that this framework achieves competitive performance on a real-world dataset, and online A/B tests demonstrate the superiority and effectiveness in improving conversion gains.

Generative retrieval (GR) has emerged as a transformative paradigm in search and recommender systems, leveraging numeric-based identifier representations to enhance efficiency and generalization. Notably, methods like TIGER, which employ Residual Quantization-based Semantic Identifiers (RQ-SID), have shown significant promise in e-commerce scenarios by effectively managing item IDs. However, a critical issue termed the " Hourglass " phenomenon, occurs in RQ-SID, where intermediate codebook tokens become overly concentrated, hindering the full utilization of generative retrieval methods. This paper analyses and addresses this problem by identifying data sparsity and long-tailed distribution as the primary causes. Through comprehensive experiments and detailed ablation studies, we analyze the impact of these factors on codebook utilization and data distribution. Our findings reveal that the “Hourglass” phenomenon substantially impacts the performance of RQ-SID in generative retrieval. We propose effective solutions to mitigate this issue, thereby significantly enhancing the effectiveness of generative retrieval in real-world E-commerce applications.

Retrieving relevant items that match users' queries from billion-scale corpus forms the core of industrial e-commerce search systems, in which embedding-based retrieval (EBR) methods are prevailing. These methods adopt a two-tower framework to learn embedding vectors for query and item separately and thus leverage efficient approximate nearest neighbor (ANN) search to retrieve relevant items. However, existing EBR methods usually ignore inconsistent user behaviors in industrial multi-stage search systems, resulting in insufficient retrieval efficiency with a low commercial return. To tackle this challenge, we propose to improve EBR methods by learning Multi-level Multi-Grained Semantic Embeddings(MMSE). We propose the multi-stage information mining to exploit the ordered, clicked, unclicked and random sampled items in practical user behavior data, and then capture query-item similarity via a post-fusion strategy. We then propose multi-grained learning objectives that integrate the retrieval loss with global comparison ability and the ranking loss with local comparison ability to generate semantic embeddings. Both experiments on a real-world billion-scale dataset and online A/B tests verify the effectiveness of MMSE in achieving significant performance improvements on metrics such as offline recall and online conversion rate (CVR).

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.

Deep semantic retrieval has achieved remarkable success in online E-commerce applications. The majority of methods aim to distinguish positive items and negative items for each query by utilizing margin loss or softmax loss. Despite their decent performance, these methods are highly sensitive to hyper-parameters, i.e., margin and temperature 𝜏 , which measure the similarity of negative pairs and affect the distribution of items in metric space. How to design and choose adaptively parameters for different pairs is still an open challenge. Recently several methods have attempted to alleviate the above problem by learning each parameter through trainable/statistical methods in the recommendation. We argue that those are not suitable for retrieval scenarios, due to the agnosticism and diversity of the queries. To fully overcome this limitation, we propose a novel adaptive metric learning method that designs a simple and universal hyper-parameter-free learning method to improve the performance of retrieval. Specifically, we first propose a method that adaptive obtains the hyper-parameters by relying on the batch similarity without fixed or extra-trainable hyper-parameters. Subsequently, we adopt a symmetric metric learning method to mitigate model collapse issues. Furthermore, the proposed method is general and sheds a highlight on other fields. Extensive experiments demonstrate our method significantly outperforms previous methods on a real-world dataset, highlighting the superiority and effectiveness of our method. This method has been successfully deployed on an online E-commerce search platform and brought substantial economic benefits.

The great success of deep learning is largely driven by training over-parameterized models on massive datasets. To avoid excessive computation, extracting and training only on the most informative subset is drawing increasing attention. Nevertheless, it is still an open question how to select such a subset on which the model trained generalizes on par with the full data. In this paper, we propose dynamic margin selection (DynaMS). DynaMS leverages the distance from candidate samples to the classification boundary to construct the subset, and the subset is dynamically updated during model training. We show that DynaMS converges with large probability, and for the first time show both in theory and practice that dynamically updating the subset can result in better generalization. To reduce the additional computation incurred by the selection, a light parameter sharing proxy (PSP) is designed. PSP is able to faithfully evaluate instances following the underlying model, which is necessary for dynamic selection. Extensive analysis and experiments demonstrate the superiority of the proposed approach in data selection against many state-of-the-art counterparts on benchmark datasets.

The great success of deep learning is largely driven by training over-parameterized models on massive datasets. To avoid excessive computation, extracting and training only on the most informative subset is drawing increasing attention. Nevertheless, it is still an open question how to select such a subset on which the model trained generalizes on par with the full data. In this paper, we propose dynamic margin selection (DynaMS). DynaMS leverages the distance from candidate samples to the classification boundary to construct the subset, and the subset is dynamically updated during model training. We show that DynaMS converges with large probability, and for the first time show both in theory and practice that dynamically updating the subset can result in better generalization over previous works. To reduce the additional computation incurred by the selection, a light parameter sharing proxy (PSP) is designed. PSP is able to faithfully evaluate instances with respect to the current model, which is necessary for dynamic selection. Extensive analysis and experiments demonstrate the superiority of the proposed approach in data selection against many state-of-the-art counterparts on benchmark datasets.

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. In this paper, we propose a novel query-aware embedding Index framework, which aligns the query and item embedding space to reduce the distance between positive pairs, thereby mixing the query and item embeddings to learn better cluster centers for product quantization. Specifically, we first propose s symmetric loss to train a better two-tower to achieve space alignment. Subsequently, we propose a mixed quantization strategy to put the query embeddings into the product quantization process for bridging the gap between queries and compressed item embeddings. Extensive experiments show that our framework significantly outperforms previous models on a real-world dataset, which demonstrates the superiority and effectiveness of the framework.

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.

Learning the user-item relevance hidden in implicit feedback data plays an important role in modern recommender systems. Neural sequential recommendation models, which formulates learning the user-item relevance as a sequential classification problem to distinguish items in future behaviors from others based on the user's historical behaviors, have attracted a lot of interest in both industry and academic due to their substantial practical value. Though achieving many practical successes, we argue that the intrinsic {\bf incompleteness} and {\bf inaccuracy} of user behaviors in implicit feedback data is ignored and conduct preliminary experiments for supporting our claims. Motivated by the observation that model-free methods like behavioral retargeting (BR) and item-based collaborative filtering (ItemCF) hit different parts of the user-item relevance compared to neural sequential recommendation models, we propose a novel model-agnostic training approach called WSLRec, which adopts a three-stage framework: pre-training, top-$k$ mining, and fine-tuning. WSLRec resolves the incompleteness problem by pre-training models on extra weak supervisions from model-free methods like BR and ItemCF, while resolves the inaccuracy problem by leveraging the top-$k$ mining to screen out reliable user-item relevance from weak supervisions for fine-tuning. Experiments on two benchmark datasets and online A/B tests verify the rationality of our claims and demonstrate the effectiveness of WSLRec.

A. Detailed Introduction of PLTs and TDMs A.1. Probabilistic Label Trees (PLTs) PLTs formulate tree modelsM(T , g) as hierarchical probability estimators for the marginal distribution p(yj |x) (Jain et al., 2016; Wydmuch et al., 2018). In PLTs, the pseudo target zn is defined as zn = I( ∑ n′∈L(n) yπ(n′) ≥ 1), which implies that zn = 1 if and only if there exists n′ ∈ C(n) such that zn′ = 1. In other words, zn = 1 implies zρ(n) = 1. As a result, for any n ∈ N , corresponding p(zn|x) can be decomposed as p(zn = 1|x) = ∏

Retrieving relevant targets from an extremely large target set under computational limits is a common challenge for information retrieval and recommendation systems. Tree models, which formulate targets as leaves of a tree with trainable node-wise scorers, have attracted a lot of interests in tackling this challenge due to their logarithmic computational complexity in both training and testing. Tree-based deep models (TDMs) and probabilistic label trees (PLTs) are two representative kinds of them. Though achieving many practical successes, existing tree models suffer from the training-testing discrepancy, where the retrieval performance deterioration caused by beam search in testing is not considered in training. This leads to an intrinsic gap between the most relevant targets and those retrieved by beam search with even the optimally trained node-wise scorers. We take a first step towards understanding and analyzing this problem theoretically, and develop the concept of Bayes optimality under beam search and calibration under beam search as general analyzing tools for this purpose. Moreover, to eliminate the discrepancy, we propose a novel algorithm for learning optimal tree models under beam search. Experiments on both synthetic and real data verify the rationality of our theoretical analysis and demonstrate the superiority of our algorithm compared to state-of-the-art methods.

It is known that the Langevin dynamics used in MCMC is the gradient flow of the KL divergence on the Wasserstein space, which helps convergence analysis and inspires recent particle-based variational inference methods (ParVIs). But no more MCMC dynamics is understood in this way. In this work, by developing novel concepts, we propose a theoretical framework that recognizes a general MCMC dynamics as the fiber-gradient Hamiltonian flow on the Wasserstein space of a fiber-Riemannian Poisson manifold. The "conservation + convergence" structure of the flow gives a clear picture on the behavior of general MCMC dynamics. The framework also enables ParVI simulation of MCMC dynamics, which enriches the ParVI family with more efficient dynamics, and also adapts ParVI advantages to MCMCs. We develop two ParVI methods for a particular MCMC dynamics and demonstrate the benefits in experiments.

Particle-based variational inference methods (ParVIs) have gained attention in the Bayesian inference literature, for their capacity to yield flexible and accurate approximations. We explore ParVIs from the perspective of Wasserstein gradient flows, and make both theoretical and practical contributions. We unify various finite-particle approximations that existing ParVIs use, and recognize that the approximation is essentially a compulsory smoothing treatment, in either of two equivalent forms. This novel understanding reveals the assumptions and relations of existing ParVIs, and also inspires new ParVIs. We propose an acceleration framework and a principled bandwidth-selection method for general ParVIs; these are based on the developed theory and leverage the geometry of the Wasserstein space. Experimental results show the improved convergence by the acceleration framework and enhanced sample accuracy by the bandwidth-selection method.

Stein variational gradient descent (SVGD) is a recently proposed particle-based Bayesian inference method, which has attracted a lot of interest due to its remarkable approximation ability and particle efficiency compared to traditional variational inference and Markov Chain Monte Carlo methods. However, we observed that particles of SVGD tend to collapse to modes of the target distribution, and this particle degeneracy phenomenon becomes more severe with higher dimensions. Our theoretical analysis finds out that there exists a negative correlation between the dimensionality and the repulsive force of SVGD which should be blamed for this phenomenon. We propose Message Passing SVGD (MP-SVGD) to solve this problem. By leveraging the conditional independence structure of probabilistic graphical models (PGMs), MP-SVGD converts the original high-dimensional global inference problem into a set of local ones over the Markov blanket with lower dimensions. Experimental results show its advantages of preventing vanishing repulsive force in high-dimensional space over SVGD, and its particle efficiency and approximation flexibility over other inference methods on graphical models.

We consider doing Bayesian inference by minimizing the KL divergence on the 2-Wasserstein space $\mathcal{P}_2$. By exploring the Riemannian structure of $\mathcal{P}_2$, we develop two inference methods by simulating the gradient flow on $\mathcal{P}_2$ via updating particles, and an acceleration method that speeds up all such particle-simulation-based inference methods. Moreover we analyze the approximation flexibility of such methods, and conceive a novel bandwidth selection method for the kernel that they use. We note that $\mathcal{P}_2$ is quite abstract and general so that our methods can make closer approximation, while it still has a rich structure that enables practical implementation. Experiments show the effectiveness of the two proposed methods and the improvement of convergence by the acceleration method.

Supervised topic models leverage label information to learn discriminative latent topic representations. As collecting a fully labeled dataset is often time-consuming, semi-supervised learning is of high interest. In this paper, we present an effective semi-supervised max-margin topic model by naturally introducing manifold posterior regularization to a regularized Bayesian topic model, named LapMedLDA. The model jointly learns latent topics and a related classifier with only a small fraction of labeled documents. To perform the approximate inference, we derive an efficient stochastic gradient MCMC method. Unlike the previous semi-supervised topic models, our model adopts a tight coupling between the generative topic model and the discriminative classifier. Extensive experiments demonstrate that such tight coupling brings significant benefits in quantitative and qualitative performance.

Stein variational gradient descent (SVGD) is a nonparametric inference method, which iteratively transports a set of randomly initialized particles to approximate a differentiable target distribution, along the direction that maximally decreases the KL divergence within a vector-valued reproducing kernel Hilbert space (RKHS). Compared to Monte Carlo methods, SVGD is particle-efficient because of the repulsive force induced by kernels. In this paper, we develop the first analysis about the high dimensional performance of SVGD and emonstrate that the repulsive force drops at least polynomially with increasing dimensions, which results in poor marginal approximation. To improve the marginal inference of SVGD, we propose Marginal SVGD (M-SVGD), which incorporates structural information described by a Markov random field (MRF) into kernels. M-SVGD inherits the particle efficiency of SVGD and can be used as a general purpose marginal inference tool for MRFs. Experimental results on grid based Markov random fields show the effectiveness of our methods.