Large-scale retrieval systems must operate under strict latency constraints while maintaining high recall. Sparse retrieval offers efficiency and interpretability, whereas dense retrieval provides stronger semantic matching. Although hybrid approaches combine both signals, their interaction is often limited, especially under intersection-based retrieval. We introduce HybridSparse, an end-to-end hybrid retrieval framework that strengthens sparse--dense interaction across modeling, training, and serving. It adopts a unified encoder with a shared backbone and jointly optimizes lexical and semantic representations through co-training. To further improve alignment, we incorporate hybrid score regularization and consistency distillation, enabling more stable and effective hybrid scoring. Experiments on public benchmarks demonstrate consistent improvements over strong sparse, dense, and hybrid baselines. In large-scale production deployment for Bing advertisement retrieval, HybridSparse delivers a +1.30% RPM gain, highlighting its practical impact.

Multimodal Attributed Graphs (MAGs) are ubiquitous in real-world applications, encompassing extensive knowledge through multimodal attributes attached to nodes (e.g., texts and images) and topological structure representing node interactions. Despite its potential to advance diverse research fields like social networks and e-commerce, MAG representation learning (MAGRL) remains underexplored due to the lack of standardized datasets and evaluation frameworks. In this paper, we first propose MAGB, a comprehensive MAG benchmark dataset, featuring curated graphs from various domains with both textual and visual attributes. Based on the MAGB dataset, we further systematically evaluate two mainstream MAGRL paradigms: GNN-as-Predictor, which integrates multimodal attributes via Graph Neural Networks (GNNs), and VLM-as-Predictor, which harnesses Vision Language Models (VLMs) for zero-shot reasoning. Extensive experiments on MAGB reveal the following critical insights: (i) Modality significances fluctuate drastically with specific domain characteristics. (ii) Multimodal embeddings can elevate the performance ceiling of GNNs. However, intrinsic biases among modalities may impede effective training, particularly in low-data scenarios. (iii) VLMs are highly effective at generating multimodal embeddings that alleviate the imbalance between textual and visual attributes. These discoveries, which illuminate the synergy between multimodal attributes and graph topologies, contribute to reliable benchmarks, paving the way for future research.

User profiling is pivotal for recommendation systems, as it transforms raw user interaction data into concise and structured representations that drive personalized recommendations. While traditional embedding-based profiles lack interpretability and adaptability, recent advances with large language models (LLMs) enable text-based profiles that are semantically richer and more transparent. However, existing methods often adhere to fixed formats that limit their ability to capture the full diversity of user behaviors. In this paper, we introduce LettinGo, a novel framework for generating diverse and adaptive user profiles. By leveraging the expressive power of LLMs and incorporating direct feedback from downstream recommendation tasks, our approach avoids the rigid constraints imposed by supervised fine-tuning (SFT). Instead, we employ Direct Preference Optimization (DPO) to align the profile generator with task-specific performance, ensuring that the profiles remain adaptive and effective. LettinGo operates in three stages: (1) exploring diverse user profiles via multiple LLMs(2) evaluating profile quality based on their impact in recommendation systems, and (3) aligning the profile generation through pairwise preference data derived from task performance. Experimental results demonstrate that our framework significantly enhances recommendation accuracy, flexibility, and contextual awareness. This work enhances profile generation as a key innovation for next-generation recommendation systems.

Owing to the unprecedented capability in semantic understanding and logical reasoning, large language models (LLMs) have shown fantastic potential in developing next-generation sequential recommender systems (RSs). However, existing LLM-based sequential RSs mostly separate index generation from sequential recommendation, leading to insufficient integration between semantic information and collaborative information. On the other hand, the neglect of user-related information hinders LLM-based sequential RSs from exploiting high-order user-item interaction patterns. In this paper, we propose the End-to-End Dual Dynamic (ED2) recommender, the first LLM-based sequential RS which adopts dual dynamic index mechanism, targeting resolving the above limitations simultaneously. The dual dynamic index mechanism can not only assembly index generation and sequential recommendation into a unified LLM-backbone pipeline, but also make it practical for LLM-based sequential recommender to take advantage of user-related information. Specifically, to facilitate the LLM comprehension ability to dual dynamic index, we propose a multigrained token regulator which constructs alignment supervision based on LLMs semantic knowledge across multiple representation granularities. Moreover, the associated user collection data and a series of novel instruction tuning tasks are specially customized to capture the high-order user-item interaction patterns. Extensive experiments on three public datasets demonstrate the superiority of ED2, achieving an average improvement of 19.62% in Hit-Rate and 21.11% in NDCG.

Knowledge distillation is often used to transfer knowledge from a strong teacher model to a relatively weak student model. Traditional methods include response-based methods and feature-based methods. Response-based methods are widely used but suffer from lower upper limits of performance due to their ignorance of intermediate signals, while feature-based methods have constraints on vocabularies, tokenizers and model architectures. In this paper, we propose a liberal feature-based distillation method (LEAD). LEAD aligns the distribution between the intermediate layers of teacher model and student model, which is effective, extendable, portable and has no requirements on vocabularies, tokenizers, or model architectures. Extensive experiments show the effectiveness of LEAD on widely-used benchmarks, including MS MARCO Passage Ranking, TREC 2019 DL Track, MS MARCO Document Ranking and TREC 2020 DL Track. Our code is available in https://github.com/microsoft/SimXNS/tree/main/LEAD.

Deploying pre-trained transformer models like BERT on downstream tasks in resource-constrained scenarios is challenging due to their high inference cost, which grows rapidly with input sequence length. In this work, we propose a constraint-aware and ranking-distilled token pruning method ToP, which selectively removes unnecessary tokens as input sequence passes through layers, allowing the model to improve online inference speed while preserving accuracy. ToP overcomes the limitation of inaccurate token importance ranking in the conventional self-attention mechanism through a ranking-distilled token distillation technique, which distills effective token rankings from the final layer of unpruned models to early layers of pruned models. Then, ToP introduces a coarse-to-fine pruning approach that automatically selects the optimal subset of transformer layers and optimizes token pruning decisions within these layers through improved L0 regularization. Extensive experiments on GLUE benchmark and SQuAD tasks demonstrate that ToP outperforms state-of-the-art token pruning and model compression methods with improved accuracy and speedups. ToP reduces the average FLOPs of BERT by 8.1X while achieving competitive accuracy on GLUE, and provides a real latency speedup of up to 7.4X on an Intel CPU. Code is available at https://github.com/microsoft/Moonlit/tree/main/ToP

Trajectory-User Linking (TUL) aiming to identify users of anonymous trajectories, has recently received increasing attention due to its wide range of applications, such as criminal investigation and personalized recommendation systems. In this paper, we propose a flexible Semi-Supervised framework for Trajectory-User Linking, namely S2TUL, which includes five components: trajectory-level graph construction, trajectory relation modeling, location-level sequential modeling, a classification layer and greedy trajectory-user relinking. The first two components are proposed to model the relationships among trajectories, in which three homogeneous graphs and two heterogeneous graphs are firstly constructed and then delivered into the graph convolutional networks for converting the discrete identities to hidden representations. Since the graph constructions are irrelevant to the corresponding users, the unlabelled trajectories can also be included in the graphs, which enables the framework to be trained in a semi-supervised way. Afterwards, the location-level sequential modeling component is designed to capture fine-grained intra-trajectory information by passing the trajectories into the sequential neural networks. Finally, these two level representations are concatenated into a classification layer to predict the user of the input trajectory. In the testing phase, a greedy trajectory-user relinking method is proposed to assure the linking results satisfy the timespan overlap constraint. We conduct extensive experiments on three public datasets with six representative competitors. The evaluation results demonstrate the effectiveness of the proposed framework.

The ubiquity of mobile devices and the accompanying deployment of sensing technologies have resulted in a massive amount of trajectory data. One important fundamental task is trajectory similarity computation, which is to determine how similar two trajectories are. To enable effective and efficient trajectory similarity computation, we propose a novel robust model, namely Contrastive Learning based Trajectory Similarity Computation (CL-TSim). Specifically, we employ a contrastive learning mechanism to learn the latent representations of trajectories and then calculate the dissimilarity between trajectories based on these representations. Compared with sequential auto-encoders that are the mainstream deep learning architectures for trajectory similarity computation, CL-TSim does not require a decoder and step-by-step reconstruction, thus improving the training efficiency significantly. Moreover, considering the non-uniform sampling rate and noisy points in trajectories, we adopt two type of augmentations, i.e., point dowm-sampling and point distorting, to enhance the robustness of the proposed model. Extensive experiments are conducted on two widely-used real-world datasets, i.e., Porto and ChengDu, which demonstrate the superior effectiveness and efficiency of the proposed model.

As the fundamental basis of sponsored search, relevance modeling measures the closeness between the input queries and the candidate ads. Conventional relevance models solely rely on the textual data, which suffer from the scarce semantic signals within the short queries. Recently, user historical click behaviors are incorporated in the format of click graphs to provide additional correlations beyond pure textual semantics, which contributes to advancing the relevance modeling performance. However, user behaviors are usually arbitrary and unpredictable, leading to the noisy and sparse graph topology. In addition, there exist other types of user behaviors besides clicks, which may also provide complementary information. In this paper, we study the novel problem of heterogeneous behavior graph learning to facilitate relevance modeling task. Our motivation lies in learning an optimal and task-relevant heterogeneous behavior graph consisting of multiple types of user behaviors. We further propose a novel HBGLR model to learn the behavior graph structure by mining the sophisticated correlations between node semantics and graph topology, and encode the textual semantics and structural heterogeneity into the learned representations. Our proposal is evaluated over real-world industry datasets, and has been mainstreamed in the Bing ads. Both offline and online experimental results demonstrate its superiority.

Embedding based retrieval (EBR) is a fundamental building block in many web applications. However, EBR in sponsored search is distinguished from other generic scenarios and technically challenging due to the need of serving multiple retrieval purposes: firstly, it has to retrieve high-relevance ads, which may exactly serve user's search intent; secondly, it needs to retrieve high-CTR ads so as to maximize the overall user clicks. In this paper, we present a novel representation learning framework Uni-Retriever developed for Bing Search, which unifies two different training modes knowledge distillation and contrastive learning to realize both required objectives. On one hand, the capability of making high-relevance retrieval is established by distilling knowledge from the "relevance teacher model''. On the other hand, the capability of making high-CTR retrieval is optimized by learning to discriminate user's clicked ads from the entire corpus. The two training modes are jointly performed as a multi-objective learning process, such that the ads of high relevance and CTR can be favored by the generated embeddings. Besides the learning strategy, we also elaborate our solution for EBR serving pipeline built upon the substantially optimized DiskANN, where massive-scale EBR can be performed with competitive time and memory efficiency, and accomplished in high-quality. We make comprehensive offline and online experiments to evaluate the proposed techniques, whose findings may provide useful insights for the future development of EBR systems. Uni-Retriever has been mainstreamed as the major retrieval path in Bing's production thanks to the notable improvements on the representation and EBR serving quality.

User-item interactions in recommendations can be naturally denoted as a user-item bipartite graph. Given the success of graph neural networks (GNNs) in graph representation learning, GNN-based Collaborative Filtering (CF) methods have been proposed to advance recommender systems. These methods often make recommendations based on the learned user and item embeddings. However, we found that they do not perform well with sparse user-item graphs which are quite common in real-world recommendations. Therefore, in this work, we introduce a novel perspective to build GNN-based CF methods for recommendations which leads to the proposed framework Localized Graph Collaborative Filtering (LGCF). One key advantage of LGCF is that it does not need to learn embeddings for each user and item, which is challenging in sparse scenarios. Alternatively, LGCF aims at encoding useful CF information into a localized graph and making recommendations based on such graph. Extensive experiments on various datasets validate the effectiveness of LGCF, especially in sparse scenarios. Furthermore, empirical results demonstrate that LGCF provides complementary information to the embedding-based CF model which can be utilized to boost recommendation performance.

Prerequisite relations among concepts are important for a wide range of educational applications, such as intelligent tutoring and curriculum planning. However, concept prerequisite relation learning is not trivial due to the sparsity of prerequisite relations. In this paper, we propose a contextual-knowledge-aware concept prerequisite relation learning approach called ConLearn. Four unique properties of the proposed approach are: (1) It transfers knowledge from large language model BERT to improve contextual representations of concepts; (2) It captures concept prerequisite transition patterns by applying graph neural network on concept prerequisite graph; (3) It is equipped with self-attention mechanism to fuse information from related concepts for target concept prerequisite relation classification; (4) No handcrafted features are used in our model, which makes our model easy to implement in downstream applications. Extensive experiments on three representative datasets demonstrate that our approach significantly outperforms the state-of-the-art methods.

Vector quantization (VQ) based ANN indexes, such as Inverted File System (IVF) and Product Quantization (PQ), have been widely applied to embedding based document retrieval thanks to the competitive time and memory efficiency. Originally, VQ is learned to minimize the reconstruction loss, i.e., the distortions between the original dense embeddings and the reconstructed embeddings after quantization. Unfortunately, such an objective is inconsistent with the goal of selecting ground-truth documents for the input query, which may cause severe loss of retrieval quality. Recent works identify such a defect, and propose to minimize the retrieval loss through contrastive learning. However, these methods intensively rely on queries with ground-truth documents, whose performance is limited by the insufficiency of labeled data. In this paper, we propose Distill-VQ, which unifies the learning of IVF and PQ within a knowledge distillation framework. In Distill-VQ, the dense embeddings are leveraged as "teachers'', which predict the query's relevance to the sampled documents. The VQ modules are treated as the "students'', which are learned to reproduce the predicted relevance, such that the reconstructed embeddings may fully preserve the retrieval result of the dense embeddings. By doing so, Distill-VQ is able to derive substantial training signals from the massive unlabeled data, which significantly contributes to the retrieval quality. We perform comprehensive explorations for the optimal conduct of knowledge distillation, which may provide useful insights for the learning of VQ based ANN index. We also experimentally show that the labeled data is no longer a necessity for high-quality vector quantization, which indicates Distill-VQ's strong applicability in practice. The evaluations are performed on MS MARCO and Natural Questions benchmarks, where Distill-VQ notably outperforms the SOTA VQ methods in Recall and MRR. Our code is avaliable at https://github.com/staoxiao/LibVQ.

Learning informative representations of users and items from the historical interactions is crucial to collaborative filtering (CF). Existing CF approaches usually model interactions solely within the Euclidean space. However, the sophisticated user-item interactions inherently present highly non-Euclidean anatomy with various types of geometric patterns (i.e., tree-likeness and cyclic structures). The Euclidean-based models may be inadequate to fully uncover the intent factors beneath such hybrid-geometry interactions. To remedy this deficiency, in this paper, we study the novel problem of Geometric Disentangled Collaborative Filtering (GDCF), which aims to reveal and disentangle the latent intent factors across multiple geometric spaces. A novel generative GDCF model is proposed to learn geometric disentangled representations by inferring the high-level concepts associated with user intentions and various geometries. Empirically, our proposal is extensively evaluated over five real-world datasets, and the experimental results demonstrate the superiority of GDCF.

Ad-hoc search calls for the selection of appropriate answers from a massive-scale corpus. Nowadays, the embedding-based retrieval (EBR) becomes a promising solution, where deep learning based document representation and ANN search techniques are allied to handle this task. However, a major challenge is that the ANN index can be too large to fit into memory, given the considerable size of answer corpus. In this work, we tackle this problem with Bi-Granular Document Representation, where the lightweight sparse embeddings are indexed and standby in memory for coarse-grained candidate search, and the heavyweight dense embeddings are hosted in disk for fine-grained post verification. For the best of retrieval accuracy, a Progressive Optimization framework is designed. The sparse embeddings are learned ahead for high-quality search of candidates. Conditioned on the candidate distribution induced by the sparse embeddings, the dense embeddings are continuously learned to optimize the discrimination of ground-truth from the shortlisted candidates. Besides, two techniques: the contrastive quantization and the locality-centric sampling are introduced for the learning of sparse and dense embeddings, which substantially contribute to their performances. Thanks to the above features, our method effectively handles massive-scale EBR with strong advantages in accuracy: with up to recall gain on million-scale corpus, and up to recall gain on billion-scale corpus. Besides, Our method is applied to a major sponsored search platform with substantial gains on revenue (), Recall () and CTR (). Our code is available at https://github.com/microsoft/BiDR.

Although many applications take subtrajectories as basic units for analysis, there is little research on the similar subtrajectory search problem aiming to return a portion of a trajectory (i.e., subtrajectory), which is the most similar to a query trajectory. We find that in some special cases, when a grid-based metric is used, this problem can be formulated as a reading comprehension problem, which has been studied extensively in the field of natural language processing (NLP). By this formulation, we can obtain faster models with better performance than existing methods. However, due to the difference between natural language and trajectory (e.g., spatial relationship), it is impossible to directly apply NLP models to this problem. Therefore, we propose a Similar Subtrajectory Search with a Graph Neural Networks framework. This framework contains four modules including a spatial-aware grid embedding module, a trajectory embedding module, a query-context trajectory fusion module, and a span prediction module. Specifically, in the spatial-aware grid embedding module, the spatial-based grid adjacency is constructed and delivered to the graph neural network to learn spatial-aware grid embedding. The trajectory embedding module aims to model the sequential information of trajectories. The purpose of the query-context trajectory fusion module is to fuse the information of the query trajectory to each grid of the context trajectories. Finally, the span prediction module aims to predict the start and the end of a subtrajectory for the context trajectory, which is the most similar to the query trajectory. We conduct comprehensive experiments on two real world datasets, where the proposed framework outperforms the state-of-the-art baselines consistently and significantly.

Accurately recommending the next point of interest (POI) has become a fundamental problem with the rapid growth of location-based social networks. However, sparse, imbalanced check-in data and diverse user check-in patterns pose severe challenges for POI recommendation tasks. Knowledge-aware models are known to be primary in leveraging these problems. However, as most knowledge graphs are constructed statically, sequential information is yet integrated. In this work, we propose a meta-learned sequential-knowledge-aware recommender (Meta-SKR), which utilizes sequential, spatio-temporal, and social knowledge to recommend the next POI for a location-based social network user. The framework mainly contains four modules. First, in the graph construction module, a novel type of knowledge graph—the sequential knowledge graph, which is sensitive to the check-in order of POIs—is built to model users’ check-in patterns. To deal with the problem of data sparsity, a meta-learning module based on latent embedding optimization is then introduced to generate user-conditioned parameters of the subsequent sequential-knowledge-aware embedding module, where representation vectors of entities (nodes) and relations (edges) are learned. In this embedding module, gated recurrent units are adapted to distill intra- and inter-sequential knowledge graph information. We also design a novel knowledge-aware attention mechanism to capture information surrounding a given node. Finally, POI recommendation is provided by inferring potential links of knowledge graphs in the prediction module. Evaluations on three real-world check-in datasets show that Meta-SKR can achieve high recommendation accuracy even with sparse data.

Human mobility recovery is of great importance for a wide range of location-based services. However, recovering human mobility is not trivial because of three challenges: 1) complex transition patterns among locations; 2) multi-level periodicity and shifting periodicity of human mobility; 3) sparsity of the collected trajectory data. In this paper, we propose PeriodicMove, a neural attention model based on graph neural network for human mobility recovery from lengthy and sparse trajectories. In PeriodicMove, we first construct a directed graph for each trajectory and capture complex location transition patterns using graph neural network. Then, we design two attention mechanisms which capture multi-level periodicity and shifting periodicity of human mobility respectively. Finally, a spatial-aware loss function is proposed to incorporate spatial proximity into the model optimization, which alleviates the data sparsity problem. We perform extensive experiments and the evaluation results demonstrate that PeriodicMove yields significant improvements over the competitors on two representative real-life mobility datasets. In addition, by providing high-quality mobility data, our model can benefit a variety of mobility-oriented downstream applications.

Sponsored search ads appear next to search results when people look for products and services on search engines. In recent years, they have become one of the most lucrative channels for marketing. As the fundamental basis of search ads, relevance modeling has attracted increasing attention due to the significant research challenges and tremendous practical value. Most existing approaches solely rely on the semantic information in the input query-ad pair, while the pure semantic information in the short ads data is not sufficient to fully identify user's search intents. Our motivation lies in incorporating the tremendous amount of unsupervised user behavior data from the historical search logs as the complementary graph to facilitate relevance modeling. In this paper, we extensively investigate how to naturally fuse the semantic textual information with the user behavior graph, and further propose three novel AdsGNN models to aggregate topological neighborhood from the perspectives of nodes, edges and tokens. Furthermore, two critical but rarely investigated problems, domain-specific pre-training and long-tail ads matching, are studied thoroughly. Empirically, we evaluate the AdsGNN models over the large industry dataset, and the experimental results of online/offline tests consistently demonstrate the superiority of our proposal.

Text encoders based on C-DSSM or transformers have demonstrated strong performance in many Natural Language Processing (NLP) tasks. Low latency variants of these models have also been developed in recent years in order to apply them in the field of sponsored search which has strict computational constraints. However these models are not the panacea to solve all the Natural Language Understanding (NLU) challenges as the pure semantic information in the data is not sufficient to fully identify the user intents. We propose the TextGNN model that naturally extends the strong twin tower structured encoders with the complementary graph information from user historical behaviors, which serves as a natural guide to help us better understand the intents and hence generate better language representations. The model inherits all the benefits of twin tower models such as C-DSSM and TwinBERT so that it can still be used in the low latency environment while achieving a significant performance gain than the strong encoder-only counterpart baseline models in both offline evaluations and online production system. In offline experiments, the model achieves a 0.14% overall increase in ROC-AUC with a 1% increased accuracy for long-tail low-frequency Ads, and in the online A/B testing, the model shows a 2.03% increase in Revenue Per Mille with a 2.32% decrease in Ad defect rate.

We propose a novel training scheme for fast matching models in Search Ads, motivated by practical challenges. The first challenge stems from the pursuit of high throughput, which prohibits the deployment of inseparable architectures, and hence greatly limits model accuracy. The second problem arises from the heavy dependency on human provided labels, which are expensive and time-consuming to collect, yet how to leverage unlabeled search log data is rarely studied. The proposed training framework targets on mitigating both issues, by treating the stronger but undeployable models as annotators, and learning a deployable model from both human provided relevance labels and weakly annotated search log data. Specifically, we first construct multiple auxiliary tasks from the enumerated relevance labels, and train the annotators by jointly learning from those related tasks. The annotation models are then used to assign scores to both labeled and unlabeled training samples. The deployable model is firstly learnt on the scored unlabeled data, and then fine-tuned on scored labeled data, by leveraging both labels and scores via minimizing the proposed label-aware weighted loss. According to our experiments, compared with the baseline that directly learns from relevance labels, training by the proposed framework outperforms it by a large margin, and improves data efficiency substantially by dispensing with 80% labeled samples. The proposed framework allows us to improve the fast matching model by learning from stronger annotators while keeping its architecture unchanged. Meanwhile, it offers a principled manner to leverage search log data in the training phase, which could effectively alleviate our dependency on human provided labels.